mirror of
https://github.com/aws/agent-toolkit-for-aws.git
synced 2026-11-22 05:18:04 +08:00
fix(aws-startup-advisor): refresh GPT-5.4/GPT-5.5 Bedrock endpoint metadata from the public model cards
The plugin recorded GPT-5.4 and GPT-5.5 as bedrock-mantle-only (no CRIS, no inference profile, no bedrock-runtime surface) with Chat Completions "unconfirmed", a 272K context window and no long-context price tier, and the canonical catalog cited "internal reference" probes as provenance. The public model cards no longer say any of that. This PR corrects the metadata only; it does not change endpoint selection, dispatch, or add new code paths. Sources (read 2026-11-06 via the AWS Documentation MCP tools): - https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html - https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html What the cards state, now recorded verbatim in the catalog and prose: - Endpoints: bedrock-runtime supported (Responses, Chat Completions, Converse; not Messages/Invoke) and bedrock-mantle supported (Responses, Chat Completions). Runtime has no in-region form; requests name us.openai.gpt-5.4 / global.openai.gpt-5.4 and us.openai.gpt-5.5 / global.openai.gpt-5.5. Mantle in-region: 5.4 = us-east-1, us-east-2, us-west-2, us-gov-west-1; 5.5 = us-east-1, us-east-2. US Geo CRIS from 6 regions, Global CRIS from 33. - Context 1,050,000 tokens shared, 128,000 max output. - Pricing (Standard, In-Region only; Priority/Flex unsupported): 5.4 short 2.75 / 0.275 cache read / 16.50, long 5.50 / 0.55 / 24.75, GovCloud short 3.375 / 0.3375 / 20.25; 5.5 short 5.50 / 0.55 / 33.00, long 11.00 / 1.10 / 49.50. No Geo/Global price rows, so prose keeps the ~10% figure for these two on every inference option and never claims parity. - Runtime features: invocation logs, streaming, Guardrails and application inference profiles (Converse only), implicit prompt caching and structured outputs (Responses only); no server-side tool use. Changes: - skills/shared/ai/models/openai-bedrock.json: GPT-5.4/5.5 rows re-sourced to the public cards (no "internal reference"), mantle_openai_chat available, runtime_converse path with CRIS, inference_profiles (us/global), region_availability, pricing tiers, context/output limits; path-level constraints corrected. - Canonical shared docs: ai-openai-to-bedrock.md (Tier 0 table, region gate, caching, Bedrock-native section, volume guidance), ai-migration-guardrails.md, schema-design-aws-ai.md (ban is on BARE ids paired with converse/runtime, not on the models), bedrock-pricing-cache.md (rows, long-context table, caching note, GovCloud rate). - gcp-to-aws and azure-to-aws: references/shared/openai-on-bedrock.md, shared/schema-design-aws-ai.md (azure), design-ai.md, design.md/generate.md _assert lines (azure), generate-artifacts-ai.md (azure), generate-ai.md, discover-preview.md and clarify-ai-only.md (gcp), clarify-ai-only.md (azure). - bedrock-migration-assessment clarify-ai-only.md; agent-advisor model-recommend.md (documents the engine's current GPT-5.x reshape as an engine limitation, not a model fact). - llm-to-bedrock: resolve-bedrock-model-id.md (Case A/B), behavior-delta openai-to-bedrock.md, SKILL.md (two id-form sentences); docstrings/comments in preflight_bedrock.py, iam_policy.py, bedrock_pricing.py and the preflight 404 remedy string. The is_mantle_model / is_mantle_gpt ID-form logic is unchanged. - agents: llm2bedrock-code-rewriter.md (Converse rule is id-form-dependent), llm2bedrock-report-generator.md (cost text), llm2bedrock-prompt-evaluator.md. - openai_model_recommendation.py: only the chat_completions_to_responses block message text (dropped the "GPT-5.x on Mantle rejects Chat Completions" claim). - Tests: tests/test_gpt6_review_round2.py, test_gpt6_review_round5.py, test_gpt6_round4_findings.py and agent-advisor scripts/test_openai_model_recommendation.py now assert the corrected public facts (CRIS ids present, "mantle-only"/"no CRIS"/"unconfirmed" absent, catalog provenance is the public card). New regression tests in test_preflight_bedrock.py, test_iam_policy.py, test_bedrock_pricing.py prove us./global.openai.gpt-5.4|5.5 classify as runtime and bare ids as mantle in all three helpers. - Vendored copies under agent-advisor, gcp-to-aws, azure-to-aws re-synced with tools/sync-vendored.py. Validation (all run from the worktree with mise Python 3.12): - mise -C plugins/aws-startup-advisor run ci: PASS - lint:md 0 errors; lint:model-id OK (19 retired IDs enforced); lint:model-staleness fresh (4d old, 45d window); lint:model-catalog-drift OK (3 pre-existing warnings: sonnet-5-5, haiku-5-5, gpt-6.1-terra not in catalog; unrelated); lint:startup-model-pricing OK; lint:artifacts PASS (99 checked, 0 errors); lint:vendored-parity PASS (62 files byte-identical); lint:skill-refs PASS (1482 checked, 0 missing); lint:consent-disclosure OK; test: 1395 passed, 222 subtests passed; test:telemetry: 311 passed; test:heroku-source-review: 44 pass, 0 fail - agent-advisor pytest (skills/agent-advisor/scripts): 564 passed, 0 failed (562 at base + 2 new) - llm-to-bedrock pytest (skills/llm-to-bedrock/scripts): 418 passed, 0 failed (415 at base + 3 new; needs --with httpx --with aws-bedrock-token-generator) - root: python3 tools/validate.py "All validations passed"; mise run lint:md 0 issues / 2763 files; lint:manifests, lint:skill-parity, lint:spec all pass; test:tools 6 OK. mise run build's security step fails on a pre-existing gitleaks false positive in upstream commitf8d09fd1(tests/test_bedrock_migration_assessment_discover.py:118, "claude-3.5-sonnet"); gitleaks protect over this change set: no leaks found. - grep over plugins/ for mantle-only / "no CRIS" / "no inference profile" / "no runtime path" / "unconfirmed" / "internal reference" / "no 1M tier" near GPT-5.4/5.5: no remaining claims (only negative test assertions and the bare-id rule "bare proprietary openai.gpt-* IDs use Mantle only"). Base:0a218f8cc1(upstream/main)
This commit is contained in:
@@ -257,7 +257,7 @@ client = anthropic.Anthropic(
|
||||
|
||||
Preserve compatible SDK request/response shapes. A model version change still requires §9 target checks, including typed response parsing for Opus 5.5. Apply the Chat Completions → Responses reshape only when that API change was selected. Preserve the exact validated runtime profile ID or ARN. After applying the changes above, skip the Converse-specific guidance in the rest of §8; the §9 behavior-delta application still applies normally.
|
||||
|
||||
**Converse for a proprietary GPT target is family-dependent.** GPT-5.5 / GPT-5.4 (`openai.gpt-5.5`, `openai.gpt-5.4`) have no `bedrock-runtime` surface — a boto3 `converse()` call against them fails at runtime; if your context pairs one with the Converse path, stop and report the contradiction. GPT-5.6 and GPT-6 (Astra, Sol, Luna, 6.1 Sol) DO have a `bedrock-runtime` path via CRIS ids (`us.`/`in.`/`global.` prefixed): when the plan says `migration_path: runtime_openai_cris`, a Converse or Responses rewrite against the CRIS id is valid (base URL `bedrock-runtime.{region}.amazonaws.com/openai/v1` for the OpenAI-compatible APIs; note Guardrails are Converse-only and prompt caching Responses-only on runtime). For a Responses-based source app, the mantle express lane above remains the default and smallest change.
|
||||
**Converse for a proprietary GPT target is id-form-dependent, not family-dependent.** A **bare** proprietary GPT id (`openai.gpt-5.5`, `openai.gpt-5.4`, `openai.gpt-5.6-*`, `openai.gpt-6-*`) is a `bedrock-mantle` id with no `bedrock-runtime` in-region form — a boto3 `converse()` call against a bare id fails at runtime; if your context pairs a bare id with the Converse path, stop and report the contradiction. GPT-5.6, GPT-6 (Astra, Sol, Luna, 6.1 Sol) **and GPT-5.5 / GPT-5.4** (model cards read 2026-11-06: `us.openai.gpt-5.5` / `global.openai.gpt-5.5`, `us.openai.gpt-5.4` / `global.openai.gpt-5.4`; Responses, Chat Completions and Converse on runtime, no Invoke) all have a `bedrock-runtime` path via CRIS ids (`us.`/`in.`/`global.` prefixed): when the plan says `migration_path: runtime_openai_cris`, a Converse, Responses or Chat Completions rewrite against the CRIS id is valid (base URL `bedrock-runtime.{region}.amazonaws.com/openai/v1` for the OpenAI-compatible APIs; note Guardrails are Converse-only and prompt caching Responses-only on runtime). For a Responses-based source app, the mantle express lane above remains the default and smallest change.
|
||||
|
||||
### Converse rewrite (default)
|
||||
|
||||
|
||||
@@ -229,12 +229,13 @@ Interpret the result:
|
||||
region. These models are **in-region only**, so there is no cross-region
|
||||
inference profile to fall back to and `resolve-bedrock-model-id`'s
|
||||
inference-profile ranking does not apply. Return
|
||||
`{ blocked: { reason: 'model_unresolvable', detail: '<the exact error> — <TARGET_MODEL_ID> is not available on the bedrock-mantle endpoint in <REGION>. Mantle is in-region only; for Astra use Virginia/Oregon Standard Mantle or a supported us./global. runtime profile with an explicit plan change; GPT-5.6 may use its us./in./global. CRIS forms, while GPT-5.5/5.4 need a supported region or a different model.' } }`.
|
||||
`{ blocked: { reason: 'model_unresolvable', detail: '<the exact error> — <TARGET_MODEL_ID> is not available on the bedrock-mantle endpoint in <REGION>. Mantle is in-region only; for Astra use Virginia/Oregon Standard Mantle or a supported us./global. runtime profile with an explicit plan change; GPT-5.6 may use its us./in./global. CRIS forms and GPT-5.5/5.4 their us./global. CRIS forms with an explicit plan change, or pick a supported mantle region or a different model.' } }`.
|
||||
Use `model_unresolvable` rather than a new reason — its recovery path (user
|
||||
picks or pastes an ID, recorded in `resolved_model_overrides`) is exactly right
|
||||
here. Astra may use only supported `us.` / `global.` CRIS candidates; check its own region matrix. For GPT-5.6 a `us.`/`in.`/`global.`-prefixed CRIS candidate is legitimate
|
||||
(it switches the run to the Converse path); for GPT-5.5 / GPT-5.4 no prefixed
|
||||
form exists — do not offer one.
|
||||
(it switches the run to the Converse path); for GPT-5.5 / GPT-5.4 the `us.`/`global.`
|
||||
prefixed forms exist on `bedrock-runtime` (model cards read 2026-11-06) and are offered
|
||||
the same way — never silently, always as an explicit plan change.
|
||||
- **HTTP 401 / 403** — distinguish the two causes from the message. If it names
|
||||
model access, return `{ blocked: { reason: 'model_access', ... } }` as in §6.
|
||||
Otherwise it is IAM: the principal needs the `bedrock-mantle:*` actions
|
||||
|
||||
@@ -143,7 +143,7 @@ The pricing script only covers Bedrock models. For the source provider (OpenAI,
|
||||
|
||||
**Astra:** use its separately dated Bedrock rates and context/inference tier. Its source price is absent from the table below; verify that price or mark the source comparison unavailable. Do not estimate Astra from a lower GPT tier or assume GPT-5.x parity/premium.
|
||||
|
||||
**GPT-5.x same-model pair cost depends on the inference option — state it conditionally, never as a flat figure.** In-Region and Geo CRIS are priced at OpenAI's _data residency_ tier (1.10x standard), so those moves cost ~10% more — quote the increase plainly and carry the non-cost justification (commitments, governance, residency, prompt caching, no behaviour delta). **GPT-5.6 on Global CRIS is priced at the standard rate — cost parity** — applicable only when the workload has no data-residency constraint; when the plan's `migration_path` is `runtime_openai_cris` with a `global.` id, report parity, not a premium. GPT-5.5/5.4 have no CRIS, so the ~10% figure always applies to them. **If a pair's workload exceeds 272K context**, price the Bedrock side at the GPT-5.6 long-context tier — 2.0x input, 1.5x output per option — and show that line separately; GPT-5.5/5.4 have no long-context tier so their usable window is 272K.
|
||||
**GPT-5.x same-model pair cost depends on the inference option — state it conditionally, never as a flat figure.** In-Region and Geo CRIS are priced at OpenAI's _data residency_ tier (1.10x standard), so those moves cost ~10% more — quote the increase plainly and carry the non-cost justification (commitments, governance, residency, prompt caching, no behaviour delta). **GPT-5.6 on Global CRIS is priced at the standard rate — cost parity** — applicable only when the workload has no data-residency constraint; when the plan's `migration_path` is `runtime_openai_cris` with a `global.` id, report parity, not a premium. GPT-5.5/5.4 do have `us.`/`global.` CRIS ids on `bedrock-runtime`, but their model cards (read 2026-11-06) publish In-Region pricing only — so the ~10% figure applies to them on every inference option until a Global rate is published; never report parity for them. **If a pair's workload exceeds 272K context**, price the Bedrock side at the long-context tier and show that line separately: GPT-5.6 is 2.0x input / 1.5x output per option; GPT-5.5 is 11.00 / 1.10 cache read / 49.50 and GPT-5.4 is 5.50 / 0.55 cache read / 24.75 per 1M tokens (In-Region, from their cards), applied to all tokens of the request, within a 1,050,000-token context window.
|
||||
|
||||
If the source model is not in this table (likely for any model released after the table's last update — check the provider's public pricing page if you know current rates), note it as a gap in the Risk Assessment section, label the figure "(estimated from `<similar model>`)", and estimate using the closest listed model's pricing.
|
||||
|
||||
|
||||
+11
-5
@@ -94,17 +94,23 @@ reuse the Anthropic `preserve_messages_api` switch):
|
||||
|
||||
OpenAI is handled by a dedicated provider module (`openai-bedrock.json` catalog).
|
||||
Recognized GPT-5.6 and GPT-6 source identities are evaluated against their exact same-model
|
||||
catalog entry before capability/numeric fallbacks. GPT-5.x on Mantle is Responses-only, so a Chat
|
||||
Completions source is reshaped rather than routed to `mantle_openai_chat`. GPT-6 Chat Completions
|
||||
sources with required or preferred API continuity instead keep that surface through
|
||||
`mantle_openai_chat`; GPT-6 Responses sources use `mantle_openai_responses`. Both Mantle paths are
|
||||
catalog entry before capability/numeric fallbacks. The engine routes only recognized GPT-6 Chat
|
||||
Completions sources to `mantle_openai_chat`; a GPT-5.x Chat Completions source is currently
|
||||
reshaped to Responses. That reshape is a current engine limitation, not a model fact: the catalog
|
||||
records Chat Completions as supported on Mantle for GPT-5.5 and GPT-5.4 (model cards read
|
||||
2026-11-06) and as unchecked for GPT-5.6 Sol. GPT-6 Chat Completions sources with required or
|
||||
preferred API continuity keep that surface through `mantle_openai_chat`; GPT-6 Responses sources
|
||||
use `mantle_openai_responses`. Both Mantle paths are
|
||||
in-region only and enforce the selected model's cataloged `region_availability`: Astra reaches
|
||||
us-east-1 and us-west-2, while Sol, Luna, and 6.1 Sol reach us-east-1 only. Outside that footprint,
|
||||
the engine skips the same-model Mantle candidate and evaluates its existing capability/numeric
|
||||
fallbacks. Governance requirements select the exact same GPT-5.6 or GPT-6 model through
|
||||
`runtime_converse` only when the requested Global or geography-scoped CRIS profile is cataloged
|
||||
for that model and source region; an unsupported profile falls through to an eligible documented
|
||||
runtime candidate or remains an explicit decision. GPT-5.5/5.4 remain mantle-only. GPT-6 Astra's
|
||||
runtime candidate or remains an explicit decision. GPT-5.5/5.4 also carry a cataloged
|
||||
`runtime_converse` path via `us.`/`global.` CRIS ids (model cards read 2026-11-06); the engine does
|
||||
not yet recognize them as same-model source identities, so a GPT-5.5/5.4 source under governance
|
||||
requirements still lands on the Claude Converse tier until that recognition is added. GPT-6 Astra's
|
||||
runtime surface lacks `Invoke` support (Responses/Chat Completions/Converse only), while
|
||||
Sol/Luna/6.1 Sol support all four APIs. Every selected path remains provisional until it is probed
|
||||
in the target account and region. Azure OpenAI remains an explicit `provider_module_pending`
|
||||
|
||||
+20
-12
@@ -133,8 +133,8 @@ See `vendored/ai/ai-model-lifecycle.md` for lifecycle details. **Do not recommen
|
||||
| GPT-5.6 Sol | openai.gpt-5.6-sol | OpenAI | 4.40 | 22.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 8.80/33.00) |
|
||||
| GPT-5.6 Terra | openai.gpt-5.6-terra | OpenAI | 2.20 | 13.20 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 4.40/19.80) |
|
||||
| GPT-5.6 Luna | openai.gpt-5.6-luna | OpenAI | 0.22 | 1.32 | 272K | fast | active (mantle in-region + runtime CRIS; 1M tier 0.44/1.98) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle only; no 1M tier) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle only; no 1M tier) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 11.00/49.50; card read 2026-11-06) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 5.50/24.75; card read 2026-11-06) |
|
||||
| Gemma 4 31B | google.gemma-4-31b | Google | 0.14 | 0.40 | 256K | mid | active (Mantle only) |
|
||||
| Gemma 4 26B-A4B | google.gemma-4-26b-a4b | Google | 0.13 | 0.40 | 256K | efficient | active (Mantle only) |
|
||||
| Gemma 4 E2B | google.gemma-4-e2b | Google | 0.04 | 0.08 | 128K | budget | active (Mantle only) |
|
||||
@@ -353,17 +353,24 @@ Regions (mantle): us-east-1 / us-east-2 for all five; us-west-2 additionally car
|
||||
| GPT-5.5 | 5.50 | 33.00 | — | 0.55 |
|
||||
| GPT-5.4 | 2.75 | 16.50 | — | 0.275 |
|
||||
|
||||
#### Long context (1M) — GPT-5.6 only, 2.0x input / 1.5x output
|
||||
#### Long context (>272K input tokens) — GPT-5.6 at 2.0x input / 1.5x output; GPT-5.5 / GPT-5.4 per their model cards
|
||||
|
||||
| Model | Input $/1M | Output $/1M |
|
||||
| ------------- | ---------- | ----------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 |
|
||||
| Model | Input $/1M | Output $/1M | Cache read $/1M |
|
||||
| ------------- | ---------- | ----------- | --------------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 | — |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 | — |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 | — |
|
||||
| GPT-5.5 | 11.00 | 49.50 | 1.10 |
|
||||
| GPT-5.4 | 5.50 | 24.75 | 0.55 |
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier. GPT-5.5 and GPT-5.4 have no such tier, so
|
||||
their usable window is 272K. Global CRIS is published at OpenAI's standard list price (cost parity, GPT-5.6 only —
|
||||
see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
A workload above 272K input tokens must be priced at the long-context tier, applied to all input and output tokens
|
||||
of the request. GPT-5.5 and GPT-5.4 rows read from their model cards on 2026-11-06
|
||||
([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html),
|
||||
[GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)); both cards state a
|
||||
1,050,000-token context window (128,000 max output) and list `us.`/`global.` CRIS ids on `bedrock-runtime`, but
|
||||
publish the In-Region inference option only. Global CRIS is published at OpenAI's standard list price (cost parity,
|
||||
GPT-5.6 only — see `openai-on-bedrock.md`). GovCloud (US-West) GPT-5.4 short-context is 3.375 / 20.25 (cache read
|
||||
0.3375) per the 2026-11-06 card; an earlier revision recorded 3.30 / 19.80.
|
||||
|
||||
> **One source conflict.** These models are absent
|
||||
> from the AWS Price List API entirely: querying `AmazonBedrock` returns only `gpt-oss` and GPT OSS Safeguard, and
|
||||
@@ -373,7 +380,8 @@ see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
|
||||
**Prompt caching (GPT-5.6 only):** cached input read at a 90% discount, cache write at 1.25x uncached input, minimum
|
||||
1,024-token prefix, up to 4 breakpoints, ≥30-minute retention. Cached input is exempt from the input-TPM quota.
|
||||
GPT-5.5 and GPT-5.4 do not list prompt caching support.
|
||||
GPT-5.5 and GPT-5.4 list only implicit prompt caching on `bedrock-runtime` Responses (automatic, no request
|
||||
parameters; cache-read rate above, no cache-write rate) and none on mantle (model cards read 2026-11-06).
|
||||
|
||||
#### Open-weight gpt-oss
|
||||
|
||||
|
||||
+70
-19
@@ -3,8 +3,8 @@
|
||||
"provider": "openai",
|
||||
"verified_at": "2026-09-24",
|
||||
"verified_region": "us-east-2",
|
||||
"source": "Opus 5.5 cross-family target: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5-5.html (retrieved 2026-09-24). OpenAI and other unchanged entries retain their evidence from the superseded 2026-08-21 catalog (removed; see git history).",
|
||||
"disclaimer": "The Opus cross-family target is updated to Opus 5.5; same-model OpenAI paths and their earlier evidence remain unchanged. This is a migration target catalog, not an account-availability claim. Verify the chosen model, API, region and CRIS profile before implementation. Opus 5.5 uses CRIS on runtime outside eu-west-2, has 1M context/128K output, always-on adaptive thinking, and no Batch tier. Fable and Mythos remain excluded from default selection. Opus 5.5 is the only Opus model in the automatic selection pool; older Opus snapshots and pricing records are historical references. The runtime in-region exception in eu-west-2 was verified 2026-10-03; the original catalog and other model evidence dates remain unchanged.",
|
||||
"source": "Opus 5.5 cross-family target: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5-5.html (retrieved 2026-09-24). GPT-5.5 and GPT-5.4 rows re-read from their public model cards on 2026-11-06 (https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html, https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html): both now list bedrock-runtime (Responses, Chat Completions, Converse via us./global. CRIS ids) and Chat Completions on bedrock-mantle. Other OpenAI entries retain their evidence from the superseded 2026-08-21 catalog (removed; see git history).",
|
||||
"disclaimer": "The Opus cross-family target is updated to Opus 5.5; the GPT-5.5 and GPT-5.4 entries were re-read from their public model cards on 2026-11-06 (runtime CRIS paths, Chat Completions on mantle, 1.05M context, long-context pricing); the other same-model OpenAI paths and their earlier evidence remain unchanged. This is a migration target catalog, not an account-availability claim. Verify the chosen model, API, region and CRIS profile before implementation. Opus 5.5 uses CRIS on runtime outside eu-west-2, has 1M context/128K output, always-on adaptive thinking, and no Batch tier. Fable and Mythos remain excluded from default selection. Opus 5.5 is the only Opus model in the automatic selection pool; older Opus snapshots and pricing records are historical references. The runtime in-region exception in eu-west-2 was verified 2026-10-03; the original catalog and other model evidence dates remain unchanged.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"sdk": "openai",
|
||||
@@ -19,8 +19,8 @@
|
||||
"GPT-5.x on Bedrock Mantle is served through the Responses API"
|
||||
],
|
||||
"constraints": [
|
||||
"Mantle itself has no cross-region form for any GPT model (in-region only); for GPT-5.6 the bedrock-runtime CRIS path covers additional regions instead (see runtime_converse)",
|
||||
"GPT-5.x on Mantle rejects Chat Completions; a Chat Completions source must reshape to Responses (reference-verified on gpt-5.4; the 5.6 cards list Chat Completions as a supported API but its mantle-side availability is unprobed — verify in the target account)",
|
||||
"Mantle itself has no cross-region form for any GPT model (in-region only); for GPT-5.6, GPT-5.5 and GPT-5.4 the bedrock-runtime CRIS path covers additional regions instead (see runtime_converse)",
|
||||
"Chat Completions support on Mantle is model-specific: the GPT-5.4 and GPT-5.5 model cards list Chat Completions as supported on bedrock-mantle (read 2026-11-06); the GPT-5.6 Sol card does not. A Chat Completions source reshapes to Responses only when the selected model's card does not list Chat Completions — verify in the target account",
|
||||
"Availability must be probed in the target account and region"
|
||||
]
|
||||
},
|
||||
@@ -36,7 +36,7 @@
|
||||
"OpenAI-compatible Chat Completions surface for models that accept it"
|
||||
],
|
||||
"constraints": [
|
||||
"Not available for the GPT-5.x models listed here: rejected on gpt-5.4 (internal reference probe) and unsupported on the gpt-5.6-sol model card",
|
||||
"Model-specific: listed as supported on bedrock-mantle by the GPT-5.4 and GPT-5.5 model cards (https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html, https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html, read 2026-11-06) and by the GPT-6 cards; unsupported on the gpt-5.6-sol model card. Uses /openai/v1/chat/completions",
|
||||
"Availability and model support must be probed in the target account and region"
|
||||
]
|
||||
},
|
||||
@@ -51,10 +51,10 @@
|
||||
"strengths": [
|
||||
"Bedrock-native multi-model request surface",
|
||||
"Supports Bedrock Guardrails (Converse API only), invocation logging, and CloudWatch",
|
||||
"For GPT-5.6 this is a SAME-MODEL governance path via CRIS (us./in./global. ids; base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs); the model cards recommend runtime for new applications, and Global CRIS is priced at OpenAI's standard rate (cost parity)"
|
||||
"For GPT-5.6, GPT-5.5 and GPT-5.4 this is a SAME-MODEL governance path via CRIS (us./in./global. ids; base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs); the model cards recommend runtime for new applications, and for GPT-5.6 Global CRIS is priced at OpenAI's standard rate (cost parity)"
|
||||
],
|
||||
"constraints": [
|
||||
"GPT-5.5 / GPT-5.4 have NO runtime path — a Converse target for them means a Bedrock-native model and a boto3 rewrite of the OpenAI integration",
|
||||
"GPT-5.5 / GPT-5.4 on runtime are CRIS-only (us. and global. ids; no in-region form) per their model cards (read 2026-11-06); the cards publish In-Region (1.10x) pricing only, so do not assume Global CRIS price parity for them",
|
||||
"GPT-5.6 on runtime is CRIS-only (no in-region form); Geo CRIS stays within the geography, Global CRIS routes anywhere — pick by data-residency posture. Runtime does not support server-side tool use, structured outputs, or application inference profiles for these models; prompt caching is Responses-only",
|
||||
"Bedrock-native models generally require a Global or geography-scoped CRIS inference profile",
|
||||
"Availability must be probed in the target account and region"
|
||||
@@ -111,19 +111,47 @@
|
||||
"generation": "reasoning",
|
||||
"version": "5.5",
|
||||
"class": "reasoning",
|
||||
"context_window": 272000,
|
||||
"output_token_ceiling": "unknown",
|
||||
"context_window": 1050000,
|
||||
"output_token_ceiling": 128000,
|
||||
"inference_profiles": {
|
||||
"us": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1"],
|
||||
"global": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1", "mx-central-1", "eu-central-1", "eu-central-2", "eu-north-1", "eu-south-1", "eu-south-2", "eu-west-1", "eu-west-2", "eu-west-3", "ap-east-2", "ap-northeast-1", "ap-northeast-2", "ap-northeast-3", "ap-south-1", "ap-south-2", "ap-southeast-1", "ap-southeast-2", "ap-southeast-3", "ap-southeast-4", "ap-southeast-5", "ap-southeast-6", "ap-southeast-7", "il-central-1", "me-central-1", "me-south-1", "af-south-1", "sa-east-1"]
|
||||
},
|
||||
"region_availability": {
|
||||
"commercial": ["us-east-1", "us-east-2"],
|
||||
"govcloud": [],
|
||||
"evidence": "bedrock-mantle in-region rows of the GPT-5.5 model card, read 2026-11-06: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html"
|
||||
},
|
||||
"pricing": {
|
||||
"source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html (read 2026-11-06); USD per 1M tokens, Standard tier; the card lists the In-Region inference option only",
|
||||
"commercial_short_context_le_272k": {"input": 5.50, "cache_read": 0.55, "output": 33.00},
|
||||
"commercial_long_context_gt_272k": {"input": 11.00, "cache_read": 1.10, "output": 49.50},
|
||||
"service_tiers": {"standard": true, "priority": false, "flex": false, "reserved": false}
|
||||
},
|
||||
"capabilities": [
|
||||
"reasoning",
|
||||
"tool_or_function_calling"
|
||||
"tool_or_function_calling",
|
||||
"image_input_vision"
|
||||
],
|
||||
"capability_source": "AWS model card 2026-07-21 (launch 2026-06-01, Active): advanced coding/agentic tasks; tool calling via the Responses API's built-in tool use. Feature-level Mantle behavior not probed — verify before relying. Re-verified 2026-08-21: still mantle-only, in-region only; no CRIS, no runtime path, no long-context tier.",
|
||||
"capability_source": "AWS model card https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html (read 2026-11-06; launch 2026-06-01, Active, EOL no sooner than 2027-06-01): text+image input, text output; 1,050,000-token context shared between input and output, 128,000 max output. Mantle lists server-side and client-side tool calling; runtime lists invocation logs, streaming, Guardrails (Converse only), application inference profiles (Converse only), implicit prompt caching (Responses only), structured outputs (Responses) and reasoning summaries; server-side tool use is unsupported on runtime. Documentation claims, not account probes — verify before relying.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": false,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (2026-07-21): bedrock-mantle Programmatic Access with Responses sample code; Geo/Global inference not supported"
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): bedrock-mantle supports Responses at /openai/v1/responses; in-region us-east-1 and us-east-2 only; Geo/Global not supported on mantle"
|
||||
},
|
||||
"mantle_openai_chat": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": false,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): Chat Completions listed as supported on bedrock-mantle at /openai/v1/chat/completions, with sample code"
|
||||
},
|
||||
"runtime_converse": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": true,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): bedrock-runtime supported with Responses, Chat Completions and Converse (Messages and Invoke unsupported); no in-region form — requests must name us.openai.gpt-5.5 (Geo) or global.openai.gpt-5.5 (Global); base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs"
|
||||
}
|
||||
}
|
||||
},
|
||||
@@ -213,27 +241,50 @@
|
||||
"generation": "reasoning",
|
||||
"version": "5.4",
|
||||
"class": "reasoning",
|
||||
"context_window": "unknown",
|
||||
"output_token_ceiling": "unknown",
|
||||
"context_window": 1050000,
|
||||
"output_token_ceiling": 128000,
|
||||
"inference_profiles": {
|
||||
"us": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1"],
|
||||
"global": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1", "mx-central-1", "eu-central-1", "eu-central-2", "eu-north-1", "eu-south-1", "eu-south-2", "eu-west-1", "eu-west-2", "eu-west-3", "ap-east-2", "ap-northeast-1", "ap-northeast-2", "ap-northeast-3", "ap-south-1", "ap-south-2", "ap-southeast-1", "ap-southeast-2", "ap-southeast-3", "ap-southeast-4", "ap-southeast-5", "ap-southeast-6", "ap-southeast-7", "il-central-1", "me-central-1", "me-south-1", "af-south-1", "sa-east-1"]
|
||||
},
|
||||
"region_availability": {
|
||||
"commercial": ["us-east-1", "us-east-2", "us-west-2"],
|
||||
"govcloud": ["us-gov-west-1"],
|
||||
"evidence": "bedrock-mantle in-region rows of the GPT-5.4 model card, read 2026-11-06: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html"
|
||||
},
|
||||
"pricing": {
|
||||
"source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html (read 2026-11-06); USD per 1M tokens, Standard tier; the card lists the In-Region inference option only",
|
||||
"commercial_short_context_le_272k": {"input": 2.75, "cache_read": 0.275, "output": 16.50},
|
||||
"commercial_long_context_gt_272k": {"input": 5.50, "cache_read": 0.55, "output": 24.75},
|
||||
"govcloud_us_west_short_context_le_272k": {"input": 3.375, "cache_read": 0.3375, "output": 20.25},
|
||||
"service_tiers": {"standard": true, "priority": false, "flex": false, "reserved": false}
|
||||
},
|
||||
"capabilities": [
|
||||
"tool_or_function_calling",
|
||||
"structured_output_json",
|
||||
"reasoning",
|
||||
"streaming"
|
||||
"streaming",
|
||||
"image_input_vision"
|
||||
],
|
||||
"capability_source": "internal reference smoke tests and its reasoning-API migration notebook (Responses path, 2026-06 probes) — the only entry with feature-level Mantle behavior probes (tools, structured output, sampling); limits not stated Re-verified 2026-08-21: still mantle-only, in-region only; no CRIS, no runtime path, no long-context tier.",
|
||||
"capability_source": "AWS model card https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html (read 2026-11-06; launch 2026-06-01, Active, EOL no sooner than 2027-06-01): text+image input, text output; 1,050,000-token context shared between input and output, 128,000 max output. Mantle lists server-side and client-side tool calling; runtime lists invocation logs, response streaming, Guardrails (Converse only), application inference profiles (Converse only), implicit prompt caching (Responses only), structured outputs (JSON Schema; Responses API) and reasoning summaries; server-side tool use is unsupported on runtime. Documentation claims, not account probes — verify before relying.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": false,
|
||||
"evidence": "internal reference smoke-test default Mantle model id; client.responses.create verified live"
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): bedrock-mantle supports Responses at /openai/v1/responses; in-region us-east-1, us-east-2, us-west-2 and us-gov-west-1; Geo/Global not supported on mantle"
|
||||
},
|
||||
"mantle_openai_chat": {
|
||||
"available": false,
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": false,
|
||||
"evidence": "internal reference verified gotcha: GPT-5.x on Mantle is Responses-only; Chat Completions rejected"
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): Chat Completions listed as supported on bedrock-mantle at /openai/v1/chat/completions, with sample code"
|
||||
},
|
||||
"runtime_converse": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": true,
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): bedrock-runtime supported with Responses, Chat Completions and Converse (Messages and Invoke unsupported); no in-region form — requests must name us.openai.gpt-5.4 (Geo) or global.openai.gpt-5.4 (Global); base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs"
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
+3
-2
@@ -1271,8 +1271,9 @@ def recommend_openai_workload(workload, region, catalog):
|
||||
(
|
||||
"The selected Astra Responses path requires reshaping the Chat source."
|
||||
if model_key == "openai_gpt_6_astra"
|
||||
else "GPT-5.x on Mantle rejects Chat Completions; the source must reshape "
|
||||
"to the Responses API."
|
||||
else "The selected Mantle Responses path requires reshaping the Chat source "
|
||||
"to the Responses API (Chat Completions support is per-model: see the "
|
||||
"selected model's mantle_openai_chat catalog entry)."
|
||||
),
|
||||
"Apply the request/response/tool/state reshape deltas before cutover.",
|
||||
)
|
||||
|
||||
+41
-1
@@ -64,18 +64,58 @@ def test_responses_source_with_continuity_recommends_mantle_responses():
|
||||
assert rec["primary_model"] == "openai.gpt-5.6-sol"
|
||||
|
||||
|
||||
def test_chat_completions_gpt5_target_selects_responses_not_chat():
|
||||
def test_chat_completions_unrecognized_gpt5_source_lands_on_sol_responses():
|
||||
# The default workload's source is `gpt-5.4`, which the engine does not yet
|
||||
# recognize as a same-model identity (see _SOURCE_MODEL_KEYS), so it lands on
|
||||
# the first catalog candidate — GPT-5.6 Sol — whose card leaves Chat
|
||||
# Completions unchecked on Mantle. The reshape here follows from THAT model's
|
||||
# card, not from a blanket "GPT-5.x on Mantle is Responses-only" rule: the
|
||||
# GPT-5.4 card itself lists Chat Completions on Mantle (see the catalog test
|
||||
# below).
|
||||
rec = _recommend(
|
||||
_workload(
|
||||
source={"api_surface": "chat_completions"},
|
||||
requirements={"api_continuity": "required"},
|
||||
)
|
||||
)
|
||||
assert rec["primary_model"] == "openai.gpt-5.6-sol"
|
||||
assert OPENAI_CATALOG["models"]["openai_gpt_5_6_sol"]["paths"]["mantle_openai_chat"]["available"] is False
|
||||
assert rec["api_path"] == "mantle_openai_responses"
|
||||
assert rec["api_path"] != "mantle_openai_chat"
|
||||
assert "chat_completions_to_responses_required" in _codes(rec["blocks"])
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("catalog_key", "model_id", "card"),
|
||||
[
|
||||
("openai_gpt_5_4", "openai.gpt-5.4", "model-card-openai-gpt-54.html"),
|
||||
("openai_gpt_5_5", "openai.gpt-5.5", "model-card-openai-gpt-55.html"),
|
||||
],
|
||||
)
|
||||
def test_catalog_gpt54_gpt55_record_public_runtime_and_chat_facts(catalog_key, model_id, card):
|
||||
# Public model cards (read 2026-11-06): both models list bedrock-runtime
|
||||
# (Responses, Chat Completions, Converse via us./global. CRIS ids; no in-region
|
||||
# form), Chat Completions on bedrock-mantle, a 1,050,000-token context window
|
||||
# and 128,000 max output. Provenance must be the public card, never an
|
||||
# internal reference.
|
||||
model = OPENAI_CATALOG["models"][catalog_key]
|
||||
paths = model["paths"]
|
||||
assert paths["mantle_openai_responses"]["available"] is True
|
||||
assert paths["mantle_openai_chat"]["available"] is True
|
||||
assert paths["runtime_converse"]["available"] is True
|
||||
assert paths["runtime_converse"]["requires_cris"] is True
|
||||
assert paths["runtime_converse"]["model_id"] == model_id
|
||||
assert f"us.{model_id}" in paths["runtime_converse"]["evidence"]
|
||||
assert f"global.{model_id}" in paths["runtime_converse"]["evidence"]
|
||||
assert set(model["inference_profiles"]) == {"us", "global"}
|
||||
assert model["context_window"] == 1050000
|
||||
assert model["output_token_ceiling"] == 128000
|
||||
assert card in model["capability_source"]
|
||||
for path in paths.values():
|
||||
assert "internal reference" not in path["evidence"]
|
||||
assert "internal reference" not in model["capability_source"]
|
||||
|
||||
|
||||
def test_chat_completions_case_includes_full_reshape_deltas():
|
||||
rec = _recommend(
|
||||
_workload(
|
||||
|
||||
+4
-3
@@ -129,9 +129,10 @@ running). → `ai_constraints.cross_cloud`. Default `"latency-acceptable"`.
|
||||
| Claude (Anthropic SDK) | Same model on Bedrock | client swap only — no model change |
|
||||
| Other / multiple / don't know | ask / infer | — |
|
||||
|
||||
**Same-model rows first.** GPT-5.6/5.5/5.4 run on Bedrock via `bedrock-mantle` (Responses,
|
||||
in-region only) — the case is AWS commitments/governance/residency, not savings, and not parity
|
||||
(~10% above OpenAI standard). For sources with no Bedrock equivalent, present a same-vendor
|
||||
**Same-model rows first.** GPT-5.6/5.5/5.4 run on Bedrock via `bedrock-mantle` (Responses or
|
||||
Chat Completions, in-region only) or via `bedrock-runtime` CRIS ids (`us.`/`global.`) — the case
|
||||
is AWS commitments/governance/residency, not savings, and not parity (~10% above OpenAI standard
|
||||
in-region; GPT-5.6 Global CRIS is at the standard rate). For sources with no Bedrock equivalent, present a same-vendor
|
||||
upgrade AND a cross-family option. → `ai_model_baseline`.
|
||||
|
||||
**After Batch 1:** interpret answers, write `preferences-draft.json` (`metadata.draft: true`,
|
||||
|
||||
+18
-10
@@ -155,7 +155,9 @@ data-residency-tier premium, while supported Global CRIS uses standard-rate pari
|
||||
Report the actual option; do not downgrade solely for the modest verified premium.
|
||||
**GPT-5.6/GPT-6 only:** above 272K context, apply 2.0x input and 1.5x output rates only when
|
||||
the selected model and inference option support that tier; the resulting gap may change
|
||||
the assessment. GPT-5.5/5.4 have no 1M tier. **Opus 5.5:** use the selected region/profile's
|
||||
the assessment. GPT-5.5/5.4 have their own long-context (>272K) rates on their model cards
|
||||
(read 2026-11-06; see `openai-on-bedrock.md` § Pricing) — use those, not the GPT-5.6
|
||||
multiplier. **Opus 5.5:** use the selected region/profile's
|
||||
flat rates from `references/vendored/ai/bedrock-pricing-cache.md` throughout the full 1M context;
|
||||
do not apply the GPT multiplier. Verify other models' tiers individually.
|
||||
|
||||
@@ -216,15 +218,21 @@ key/token provider), model ID, and IAM (`bedrock-mantle:*`) change. Read
|
||||
`references/shared/openai-on-bedrock.md` for exact values. **A Chat Completions source is a
|
||||
supported near-drop-in for GPT-6 and GPT-5.6** (both list Chat Completions on mantle and runtime
|
||||
CRIS per their model cards) — keep `chat.completions.create` unless a Responses-only feature or
|
||||
demonstrated constraint drives a reshape. **GPT-5.5 and GPT-5.4 are the exception**: Chat
|
||||
Completions support is unconfirmed on their model cards, so a Chat Completions source on either
|
||||
still needs a planned reshape to `responses.create`. **No Converse fallback for GPT-5.5/GPT-5.4 — GPT-5.6 and
|
||||
GPT-6 differ.** GPT-5.5 and GPT-5.4 are mantle-only, in-region only: needing Guardrails/Knowledge
|
||||
Bases/logging/an unsupported region forces a `converse` path with `model_change: true` to a
|
||||
Bedrock-native model (or `gpt-oss`). GPT-5.6 Sol/Terra/Luna and all five GPT-6 models (Astra, Sol,
|
||||
Luna, 6.1 Sol, 6.1 Terra) DO have a `bedrock-runtime` CRIS path (`us.`/`global.` prefixed IDs) — for these
|
||||
models the same need is satisfied by switching endpoint, not model: record
|
||||
`migration_path: "runtime_openai_cris"` with `model_change: false`. Note: `runtime_openai_cris`
|
||||
demonstrated constraint drives a reshape. **GPT-5.5 and GPT-5.4 follow the same rule**: their
|
||||
model cards (read 2026-11-06) list Chat Completions as supported on both `bedrock-mantle` and
|
||||
`bedrock-runtime`, so a Chat Completions source on either may keep `chat.completions.create`; the
|
||||
reshape to `responses.create` is optional for them too. Plan a reshape only when the selected
|
||||
model's card leaves Chat Completions unchecked on the selected endpoint. **Every proprietary GPT
|
||||
model on Bedrock has a `bedrock-runtime` CRIS path — including GPT-5.5 and GPT-5.4.** GPT-5.6
|
||||
Sol/Terra/Luna, all five GPT-6 models (Astra, Sol, Luna, 6.1 Sol, 6.1 Terra), and GPT-5.5 /
|
||||
GPT-5.4 (`us.openai.gpt-5.5` / `global.openai.gpt-5.5`, `us.openai.gpt-5.4` /
|
||||
`global.openai.gpt-5.4`; Responses, Chat Completions and Converse on runtime, no runtime
|
||||
in-region form) have a `bedrock-runtime` CRIS path (`us.`/`global.` prefixed IDs) — for all of
|
||||
them a Guardrails/logging/application-inference-profile/unsupported-region need is satisfied by
|
||||
switching endpoint, not model: record `migration_path: "runtime_openai_cris"` with
|
||||
`model_change: false`. Reserve `converse` with `model_change: true` to a Bedrock-native model (or
|
||||
`gpt-oss`) for a need the same-model CRIS path cannot serve (Knowledge Bases, intelligent prompt
|
||||
routing). Note: `runtime_openai_cris`
|
||||
for GPT-6 Astra is valid only via Responses/Chat Completions/Converse — Astra does not support
|
||||
`Invoke`, so verify per-feature against Astra's model card before assuming parity with
|
||||
Sol/Luna/6.1 Sol. **When Guardrails is the driving need and the design keeps
|
||||
|
||||
@@ -82,7 +82,7 @@ _postconditions:
|
||||
_on_failure: _halt_and_inform
|
||||
- _assert: "WHEN aws-design-ai.json exists: every design_blocks[] row has exactly one of target_bedrock_model / target_aws_service non-null (XOR), and every document_extraction/image_analysis/speech_transcription row has target_bedrock_model null, a non-null target_aws_service, and honest_assessment 'not_applicable' — per references/shared/schema-design-aws-ai.md"
|
||||
_on_failure: _halt_and_inform
|
||||
- _assert: "WHEN aws-design-ai.json exists AND summary.ai_source is azure_openai or openai: every source model that is available on Bedrock and carried by the target region maps to itself with model_change false — not to a Claude/Nova substitute; and no proprietary GPT-5.5 or GPT-5.4 model ID is paired with a converse/bedrock-runtime migration path (bare proprietary openai.gpt-* IDs use Mantle only; runtime_openai_cris accepts documented Astra us./global. and GPT-5.6/GPT-6 profiles with caller-region checks and the selected API; GPT-5.6 and GPT-6 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation)"
|
||||
- _assert: "WHEN aws-design-ai.json exists AND summary.ai_source is azure_openai or openai: every source model that is available on Bedrock and carried by the target region maps to itself with model_change false — not to a Claude/Nova substitute; and no bare proprietary openai.gpt-* model ID (including bare openai.gpt-5.5 or openai.gpt-5.4) is paired with a converse/bedrock-runtime migration path (bare proprietary openai.gpt-* IDs use Mantle only; runtime_openai_cris accepts documented Astra us./global., GPT-5.6/GPT-6 and GPT-5.5/GPT-5.4 us./global. profiles with caller-region checks and the selected API; GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation)"
|
||||
_on_failure: _halt_and_inform
|
||||
- _assert: "WHEN aws-design-ai.json exists: regional_warnings[] and multi_model_warnings[] are present (empty is fine); agentic_design is present only when agentic_profile.is_agentic is true; and App Runner appears nowhere"
|
||||
_on_failure: _halt_and_inform
|
||||
|
||||
+4
-3
@@ -155,9 +155,10 @@ is AWS-only and ports from gcp unchanged.
|
||||
- [ ] No secret VALUE in any emitted file; credentials are env-var references.
|
||||
- [ ] Bare proprietary GPT ids use Mantle; Astra runtime calls retain the validated `us.` /
|
||||
`global.` CRIS id. Chat/Responses and runtime artifact selection match the saved design.
|
||||
- [ ] No **GPT-5.5 or GPT-5.4** model ID paired with a Converse/`bedrock-runtime` path (those two
|
||||
models only — mantle-only, no CRIS). GPT-5.6 and GPT-6 models MAY use
|
||||
`migration_path: "runtime_openai_cris"` on `bedrock-runtime` — this is NOT a violation.
|
||||
- [ ] No **bare** proprietary GPT model ID (including bare `openai.gpt-5.5` / `openai.gpt-5.4`)
|
||||
paired with a Converse/`bedrock-runtime` path. GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 MAY use
|
||||
`migration_path: "runtime_openai_cris"` on `bedrock-runtime` with a `us.`/`in.`/`global.`
|
||||
CRIS id — this is NOT a violation.
|
||||
- [ ] Plugin attribution strings say "Azure-to-AWS".
|
||||
- [ ] `generation-ai.json` validates; `rollback_plan.mechanism == "feature_flag"`.
|
||||
|
||||
|
||||
+1
-1
@@ -73,7 +73,7 @@ _postconditions:
|
||||
_on_failure: _halt_and_inform
|
||||
- _assert: "no secret VALUE from the inventory appears in any generated artifact; secrets are emitted as Secrets Manager references"
|
||||
_on_failure: _halt_and_inform
|
||||
- _assert: "WHEN aws-design-ai.json exists AND run_mode is decide_and_execute: generation-ai.json exists and validates, its rollback_plan.mechanism is 'feature_flag' with flag_name 'AI_PROVIDER' and default_value 'azure_openai'; an ai-migration/ directory was produced (setup_bedrock.sh, test_comparison.py, bedrock_monitoring.tf on every path; migrate_to_mantle.sh for the mantle path OR provider_adapter.* for the direct/gpt-oss path, not both); and no proprietary GPT-5.5 or GPT-5.4 model ID is paired with a converse/bedrock-runtime path (bare proprietary openai.gpt-* IDs use Mantle only; runtime_openai_cris accepts documented Astra us./global. and GPT-5.6/GPT-6 profiles with caller-region checks and the selected API; GPT-5.6 and GPT-6 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation). When aws-design-ai.json is absent this is vacuously satisfied"
|
||||
- _assert: "WHEN aws-design-ai.json exists AND run_mode is decide_and_execute: generation-ai.json exists and validates, its rollback_plan.mechanism is 'feature_flag' with flag_name 'AI_PROVIDER' and default_value 'azure_openai'; an ai-migration/ directory was produced (setup_bedrock.sh, test_comparison.py, bedrock_monitoring.tf on every path; migrate_to_mantle.sh for the mantle path OR provider_adapter.* for the direct/gpt-oss path, not both); and no bare proprietary openai.gpt-* model ID (including bare openai.gpt-5.5 or openai.gpt-5.4) is paired with a converse/bedrock-runtime path (bare proprietary openai.gpt-* IDs use Mantle only; runtime_openai_cris accepts documented Astra us./global., GPT-5.6/GPT-6 and GPT-5.5/GPT-5.4 us./global. profiles with caller-region checks and the selected API; GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation). When aws-design-ai.json is absent this is vacuously satisfied"
|
||||
_on_failure: _halt_and_inform
|
||||
_forbids_files:
|
||||
- azure-resource-inventory.json
|
||||
|
||||
+62
-29
@@ -2,6 +2,7 @@
|
||||
|
||||
**Last verified:** 2026-10-10
|
||||
**GPT-6.1 Terra verified:** 2026-10-27
|
||||
**GPT-5.5 / GPT-5.4 re-verified:** 2026-11-06 against their model cards ([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) — both now list `bedrock-runtime` (Responses, Chat Completions, Converse via `us.`/`global.` CRIS ids), Chat Completions on mantle, a 1.05M context window and a long-context price tier
|
||||
**Sources:** [OpenAI model cards](https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html) (per-model
|
||||
cards linked below), [GPT-5.6 launch post](https://aws.amazon.com/blogs/machine-learning/get-started-with-openai-gpt-5-6-sol-terra-and-luna-on-amazon-bedrock/),
|
||||
[GPT-5.6 GA announcement](https://aws.amazon.com/about-aws/whats-new/2026/07/openai-gpt-sol-terra/),
|
||||
@@ -32,8 +33,8 @@ risk) all defer to it. Do not restate model IDs, regions, or endpoint paths else
|
||||
| GPT-5.6 Sol | `openai.gpt-5.6-sol` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html) |
|
||||
| GPT-5.6 Terra | `openai.gpt-5.6-terra` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-terra.html) |
|
||||
| GPT-5.6 Luna | `openai.gpt-5.6-luna` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-luna.html) |
|
||||
| GPT-5.5 | `openai.gpt-5.5` | Jun 1, 2026 | 272K | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html) |
|
||||
| GPT-5.4 | `openai.gpt-5.4` | Jun 1, 2026 | 272K | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html) |
|
||||
| GPT-5.5 | `openai.gpt-5.5` | Jun 1, 2026 | 1.05M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html) |
|
||||
| GPT-5.4 | `openai.gpt-5.4` | Jun 1, 2026 | 1.05M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html) |
|
||||
| gpt-oss-120b | `openai.gpt-oss-120b` | Aug 5, 2025 | 128K | Active | open-weight; also on `bedrock-runtime` as `openai.gpt-oss-120b-1:0` |
|
||||
| gpt-oss-20b | `openai.gpt-oss-20b` | Aug 5, 2025 | 128K | Active | open-weight; also on `bedrock-runtime` as `openai.gpt-oss-20b-1:0` |
|
||||
| GPT OSS Safeguard | `openai.gpt-oss-safeguard-120b` / `-20b` | — | — | Active | content-moderation / guardrail enforcement, not general chat |
|
||||
@@ -47,8 +48,10 @@ share a generation with a future `Sol`. **GPT-6** carries forward `Sol` and `Lun
|
||||
discount) — not a new tier name (cite [openai.com/index/introducing-gpt-6-1-sol/](https://openai.com/index/introducing-gpt-6-1-sol/)).
|
||||
|
||||
> **Context-window conflict (resolved):** the GPT-5.6 launch blog states 272K for all three variants; all three
|
||||
> model cards state 1M. **The model cards are authoritative** — use 1M for GPT-5.6. GPT-5.5 and GPT-5.4 are 272K on
|
||||
> both sources. Re-check on refresh; if AWS corrects the blog, the cards still win.
|
||||
> model cards state 1M. **The model cards are authoritative** — use 1M for GPT-5.6. The GPT-5.5 and GPT-5.4 cards
|
||||
> (read 2026-11-06) state 1,050,000 tokens shared between input and output with 128,000 max output; the 272K figure
|
||||
> recorded earlier is now only their short-/long-context pricing breakpoint. Re-check on refresh; if AWS corrects
|
||||
> the blog, the cards still win.
|
||||
|
||||
**Not on Bedrock (as of this refresh):** GPT-4o, GPT-4.1, GPT-4 / GPT-4 Turbo, GPT-3.5 Turbo, the o-series
|
||||
(o1/o3/o4-mini), GPT-5 / GPT-5.1 / GPT-5.2, and the `*-Pro` variants (GPT-5.5 Pro, GPT-5.4 Pro). Sources whose model
|
||||
@@ -112,8 +115,12 @@ input and text output. The launch announcement rounds the input context to 1M; u
|
||||
exact limits for capacity checks. Browser/computer-use capability does not establish parity with
|
||||
OpenAI-hosted tools; probe the application's specific tool flow.
|
||||
|
||||
**GPT-5.5 and GPT-5.4 are `bedrock-mantle`-only and in-region only.** Their model cards list a single
|
||||
Programmatic Access row (`bedrock-mantle`, Geo/Global "Not supported") and In-Region pricing only.
|
||||
**GPT-5.5 and GPT-5.4 have TWO endpoints** (re-verified 2026-11-06 against their model cards; earlier revisions of
|
||||
this file recorded them as `bedrock-mantle`-only). Both Programmatic Access tables now carry a `bedrock-runtime`
|
||||
row (In-Region "Not supported"; Geo `us.openai.gpt-5.5` / `us.openai.gpt-5.4`; Global `global.openai.gpt-5.5` /
|
||||
`global.openai.gpt-5.4`) and a `bedrock-mantle` row (bare id, in-region only). Supported APIs: `Responses`,
|
||||
`Chat Completions` and `Converse` on `bedrock-runtime`; `Responses` and `Chat Completions` on `bedrock-mantle`;
|
||||
`Messages` and `Invoke` on neither. The pricing tables still publish the In-Region inference option only.
|
||||
|
||||
**GPT-5.6 Sol / Terra / Luna have TWO endpoints** (verified 2026-08-21 — this changed after this file's original
|
||||
2026-08-10 verification; the Refresh Checklist predicted it):
|
||||
@@ -132,13 +139,16 @@ Constraints that still break naive assumptions:
|
||||
mantle path explicitly, and the runtime example gives `bedrock-runtime.{region}.amazonaws.com/openai/v1`.
|
||||
2. **The endpoints take different model-id forms.** Mantle takes the bare `openai.gpt-5.6-*` id and has no CRIS;
|
||||
`bedrock-runtime` takes ONLY a CRIS inference-profile id (`us.` / `in.` / `global.` prefixed) and has no
|
||||
in-region form. `bedrock:ListInferenceProfiles` returns the 5.6 CRIS profiles; it never returns the bare mantle
|
||||
ids, and it returns nothing for GPT-5.5 / GPT-5.4.
|
||||
in-region form. `bedrock:ListInferenceProfiles` returns the CRIS profiles (5.6, and `us.`/`global.` for
|
||||
GPT-5.5 / GPT-5.4 per their cards); it never returns the bare mantle ids.
|
||||
3. **API surfaces differ by endpoint.** The 5.6 cards list `Responses`, `Chat Completions`, `Invoke`, and
|
||||
`Converse` as supported APIs, with runtime-side feature splits: **Guardrails are Converse-only; prompt caching is
|
||||
Responses-only on runtime**; server-side tool use, structured outputs, and application inference profiles are NOT
|
||||
supported on runtime (server-side tool calling IS supported on mantle). For GPT-5.5 / GPT-5.4 treat Responses on
|
||||
mantle as the only verified surface.
|
||||
supported on runtime (server-side tool calling IS supported on mantle). The GPT-5.5 / GPT-5.4 cards (read
|
||||
2026-11-06) list `Responses`, `Chat Completions` and `Converse` on runtime (no `Invoke`) and `Responses` and
|
||||
`Chat Completions` on mantle; on runtime they list Guardrails and application inference profiles as
|
||||
Converse-only, and implicit prompt caching and structured outputs as Responses-only. Chat Completions is a
|
||||
documented surface for GPT-5.5 / GPT-5.4 on both endpoints — probe it in the target account like any other.
|
||||
|
||||
### GPT-6 family: Ultrafast tier
|
||||
|
||||
@@ -189,8 +199,10 @@ that expires within 12 hours and is not refreshed. Do not recommend it for produ
|
||||
|
||||
**IAM:** on the mantle path, the managed policy `AmazonBedrockMantleInferenceAccess` grants what inference needs,
|
||||
including `bedrock-mantle:CreateInference` and `bedrock-mantle:CallWithBearerToken` — `bedrock:InvokeModel` does not
|
||||
authorize mantle calls. On the GPT-5.6 `bedrock-runtime` path the usual `bedrock:InvokeModel*` against the CRIS
|
||||
inference-profile ARN applies, as for any other runtime model.
|
||||
authorize mantle calls. On the GPT-5.6 / GPT-5.5 / GPT-5.4 `bedrock-runtime` path the usual `bedrock:InvokeModel*`
|
||||
against the CRIS inference-profile ARN applies, as for any other runtime model; the GPT-5.5 / GPT-5.4 cards add
|
||||
that the identity also needs `bedrock:InvokeModel` on the account's default project
|
||||
(`arn:aws:bedrock:{region}:{account-id}:project/default`).
|
||||
|
||||
**Reasoning effort:** the five GPT-5.x models accept `none`, `low`, `medium`, `high`, `xhigh`, `max`. Because these models reason
|
||||
before responding, the model's output items (which may include reasoning items) must be passed back in the next
|
||||
@@ -200,7 +212,8 @@ request for multi-turn and tool-calling flows.
|
||||
|
||||
## Regional Availability — Endpoint-Aware
|
||||
|
||||
**The mantle in-region matrix** (the only reach for GPT-5.5 / GPT-5.4, and the in-region option for GPT-5.6):
|
||||
**The mantle in-region matrix** (the in-region option for every proprietary GPT model; GPT-5.5 / GPT-5.4 rows
|
||||
re-verified 2026-11-06):
|
||||
|
||||
| Model | us-east-1 | us-east-2 | us-west-2 | us-gov-west-1 | us-gov-east-1 |
|
||||
| ------------- | --------- | --------- | --------- | ------------- | ------------- |
|
||||
@@ -230,8 +243,16 @@ any GPT-6 model, it just means using the runtime endpoint with a CRIS id.
|
||||
**GPT-5.6 additionally reaches most commercial regions via `bedrock-runtime` CRIS** (Geo `us.` / `in.`, Global
|
||||
`global.` inference profiles; the Sol card's runtime footprint spans 30+ regions). So a region outside the mantle
|
||||
matrix does NOT make the same-model path unavailable for GPT-5.6 — it means using the runtime endpoint with a CRIS
|
||||
id, with the data-residency implications of cross-region routing. For GPT-5.5 / GPT-5.4 the mantle matrix is a hard
|
||||
gate: no CRIS, no fallback.
|
||||
id, with the data-residency implications of cross-region routing.
|
||||
|
||||
**GPT-5.5 / GPT-5.4 runtime availability** (model cards read 2026-11-06): US Geo CRIS (`us.openai.gpt-5.5` /
|
||||
`us.openai.gpt-5.4`) can be called from `us-east-1`, `us-east-2`, `us-west-1`, `us-west-2`, `ca-central-1` and
|
||||
`ca-west-1`. Global CRIS (`global.openai.gpt-5.5` / `global.openai.gpt-5.4`) is listed in those six plus
|
||||
`mx-central-1`, every EU region (`eu-central-1/2`, `eu-north-1`, `eu-south-1/2`, `eu-west-1/2/3`), `ap-east-2`,
|
||||
`ap-northeast-1/2/3`, `ap-south-1/2`, `ap-southeast-1` through `-7`, `il-central-1`, `me-central-1`, `me-south-1`,
|
||||
`af-south-1` and `sa-east-1` (33 regions). No runtime in-region form, no EU/India Geo profile, and no GovCloud
|
||||
runtime row. So the mantle matrix is no longer a hard gate for GPT-5.5 / GPT-5.4; only a strictly in-region
|
||||
requirement outside the mantle matrix blocks the same-model path.
|
||||
|
||||
Verify current footprints per model card / `aws___get_regional_availability` (AWS MCP Server) — the CRIS lists move faster than this file.
|
||||
|
||||
@@ -247,7 +268,8 @@ inference-option dimension:**
|
||||
workload has no data-residency constraint.
|
||||
|
||||
So the honest cost statement is conditional, not flat: a same-model GPT-5.6 move on Global CRIS is
|
||||
**cost-neutral**; the same move in-region or Geo (and any GPT-5.5 / GPT-5.4 move) is **~10% more expensive**. The
|
||||
**cost-neutral**; the same move in-region or Geo is **~10% more expensive**. The GPT-5.5 / GPT-5.4 cards publish
|
||||
In-Region pricing only (no Geo/Global rows as of 2026-11-06), so quote ~10% for them on every inference option. The
|
||||
same Global-CRIS-is-cost-neutral / in-Region-and-Geo-is-~10%-more pattern applies to the GPT-6 family below.
|
||||
Never state either number without stating the inference option it belongs to.
|
||||
|
||||
@@ -302,12 +324,15 @@ provider rate before claiming savings or parity for the selected context tier.
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier.
|
||||
|
||||
### GPT-5.5 / GPT-5.4 — In-Region only (no CRIS, no long-context tier; usable window 272K)
|
||||
### GPT-5.5 / GPT-5.4 — In-Region pricing only; short- and long-context tiers (model cards read 2026-11-06)
|
||||
|
||||
| Model | In-Region (in · out) | Cache read | Notes |
|
||||
| ------- | -------------------- | ---------- | -------------------------------- |
|
||||
| GPT-5.5 | 5.50 · 33.00 | 0.55 | no cache-write rate published |
|
||||
| GPT-5.4 | 2.75 · 16.50 | 0.275 | GovCloud (US-West): 3.30 · 19.80 |
|
||||
The cards publish only the In-Region inference option (the `us.` / `global.` runtime CRIS ids exist but carry no
|
||||
separate price row). Long-context rates apply to all input and output tokens of a request above 272K input tokens.
|
||||
|
||||
| Model | Short context ≤272K (in · cache read · out) | Long context >272K (in · cache read · out) | Notes |
|
||||
| ------- | ------------------------------------------- | ------------------------------------------ | ---------------------------------------- |
|
||||
| GPT-5.5 | 5.50 · 0.55 · 33.00 | 11.00 · 1.10 · 49.50 | no cache-write rate published |
|
||||
| GPT-5.4 | 2.75 · 0.275 · 16.50 | 5.50 · 0.55 · 24.75 | GovCloud (US-West) short: 3.375 · 0.3375 · 20.25 |
|
||||
|
||||
> **The Luna "blog discrepancy" resolved differently than first recorded.** The AWS News Blog's 0.20 / 1.20 is not
|
||||
> an error — it is the **Global CRIS** rate, now published on the Luna card. An earlier revision of this file said
|
||||
@@ -323,8 +348,10 @@ Astra has separate Responses-only caching evidence above. Do not inherit GPT-5.6
|
||||
|
||||
Listed as a supported feature on the GPT-5.6 Sol/Terra/Luna model cards and on all five GPT-6 model cards (Astra,
|
||||
Sol, Luna, 6.1 Sol, 6.1 Terra — confirmed by the cache-write/cache-read columns in the GPT-6 pricing tables above). The
|
||||
GPT-5.5 and GPT-5.4 cards list client-side tool calling in that slot instead and do **not** list prompt caching.
|
||||
Do not assume caching on 5.5/5.4.
|
||||
GPT-5.5 and GPT-5.4 cards (read 2026-11-06) list **implicit prompt caching on `bedrock-runtime` Responses only**
|
||||
(automatic, no request parameters) with a cache-read rate and no cache-write rate; explicit caching is unsupported
|
||||
and the mantle feature list carries client-side tool calling in that slot instead. Do not assume caching on
|
||||
5.5/5.4 mantle, and do not inherit the GPT-5.6 prefix/breakpoint rules below for them.
|
||||
|
||||
| Property | Value |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
@@ -352,10 +379,10 @@ This corrects two claims that were previously applied to all Mantle traffic in t
|
||||
|
||||
- There is **no shared 10,000 RPM account limit** governing these models — the quota dimension is TPM, per model,
|
||||
per region.
|
||||
- "Switch to `bedrock-runtime`" is **not a throughput remedy for GPT-5.5 / GPT-5.4** — they have no
|
||||
`bedrock-runtime` path at all. For GPT-5.6 the runtime path DOES exist (CRIS only; see Access Paths above), so
|
||||
moving there is a legitimate option — but treat it as an endpoint/architecture choice with its own quota family
|
||||
and residency implications, not a free throughput escape hatch.
|
||||
- "Switch to `bedrock-runtime`" is **not a free throughput escape hatch**. For GPT-5.6, GPT-5.5 and GPT-5.4 the
|
||||
runtime path exists (CRIS only; see Access Paths above — GPT-5.5 / GPT-5.4 re-verified 2026-11-06), so moving
|
||||
there is a legitimate option — but treat it as an endpoint/architecture choice with its own quota family and
|
||||
residency implications.
|
||||
|
||||
The supported mitigations are: exponential backoff with a bounded retry count (`max_retries` on the OpenAI SDK),
|
||||
spreading load across minutes rather than bursting, ramping request rate gradually, and prompt caching (cached input
|
||||
@@ -396,8 +423,8 @@ This model family is moving fast (two GA waves and a repricing inside 10 weeks).
|
||||
|
||||
1. Re-read the [OpenAI model card index](https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html)
|
||||
for models added or removed, and each per-model card for lifecycle state and EOL date.
|
||||
2. Recheck the region matrix AND the GPT-5.6 CRIS footprints — for 5.5/5.4 a region change is migration-blocking;
|
||||
for 5.6 the runtime/CRIS path usually covers the region instead.
|
||||
2. Recheck the region matrix AND the CRIS footprints for every proprietary GPT model — the runtime/CRIS path
|
||||
usually covers a region outside the mantle matrix; only a strictly in-region requirement there is blocking.
|
||||
3. Recheck rates on the Bedrock pricing page OpenAI tab, and resolve any row still marked _unverified_.
|
||||
4. Recheck whether the Price List API has gained GPT-5.x coverage; if it has, drop the caveat above and let
|
||||
`estimate-ai.md` price these models from the MCP.
|
||||
@@ -412,3 +439,9 @@ This model family is moving fast (two GA waves and a repricing inside 10 weeks).
|
||||
2026), and GPT-6.1 Sol (GA Sep 29, 2026) are all confirmed on Bedrock with their own model cards and pricing
|
||||
tables. New discovery this refresh: a 6x-Standard **Ultrafast** service tier on Astra and 6.1 Sol only (Sol and
|
||||
Luna remain Standard-only) — the first new tier shape since this file started tracking the family.
|
||||
9. **2026-11-06 refresh — GPT-5.5 / GPT-5.4 re-read.** Item 5 fired again: both cards now list `bedrock-runtime`
|
||||
(Responses, Chat Completions, Converse via `us.` / `global.` CRIS ids; no in-region runtime form), Chat
|
||||
Completions on `bedrock-mantle`, a 1,050,000-token context window with 128,000 max output, a long-context
|
||||
(>272K) price tier, implicit prompt caching on runtime Responses, and a GovCloud (US-West) GPT-5.4 rate of
|
||||
3.375 · 20.25. Earlier "mantle-only / no CRIS / 272K / no long-context tier" claims for these two models were
|
||||
stale and are corrected throughout this file.
|
||||
|
||||
+17
-13
@@ -123,17 +123,20 @@ Traditional-AI capabilities (AWS AI service, not Bedrock): `document_extraction`
|
||||
|
||||
## migration_path vocabulary
|
||||
|
||||
`mantle_openai_chat` preserves a selected GPT-6 Chat Completions API on Mantle.
|
||||
`mantle_openai_chat` preserves a selected Chat Completions API on Mantle for a model whose card
|
||||
lists it there (GPT-6, GPT-5.5, GPT-5.4).
|
||||
`mantle_openai_responses` (Azure OpenAI / OpenAI source, model on Bedrock, region carries it —
|
||||
keep the SDK, `model_change: false`) · `runtime_openai_cris` (GPT-5.6/GPT-6 same-model move
|
||||
served via `bedrock-runtime` CRIS, `us.`/`global.` prefixed, `model_change: false`) · `converse`
|
||||
(Bedrock-native Converse/`bedrock-runtime`; set `model_change: true` when **GPT-5.5 or GPT-5.4
|
||||
specifically** must move off mantle for Guardrails / Knowledge Bases / logging / an unsupported
|
||||
region, or when the GPT-6/5.6 CRIS path itself cannot serve the need) · `gpt-oss` (OpenAI-lineage
|
||||
model on the Bedrock-native runtime) · `direct` (framework-agnostic Bedrock SDK swap). **GPT-5.5
|
||||
and GPT-5.4** are NEVER paired with a `converse`/`bedrock-runtime` path — those two models only
|
||||
are mantle-only, no CRIS. GPT-5.6 and GPT-6 models MAY use `runtime_openai_cris` on
|
||||
`bedrock-runtime` — this is not a violation. Note: `runtime_openai_cris` for GPT-6 Astra is valid
|
||||
keep the SDK, `model_change: false`) · `runtime_openai_cris` (GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4
|
||||
same-model move served via `bedrock-runtime` CRIS, `us.`/`in.`/`global.` prefixed,
|
||||
`model_change: false`) · `converse` (Bedrock-native Converse/`bedrock-runtime`; set
|
||||
`model_change: true` when the workload needs Knowledge Bases, intelligent prompt routing, or
|
||||
another feature the same-model CRIS path itself cannot serve) · `gpt-oss` (OpenAI-lineage model
|
||||
on the Bedrock-native runtime) · `direct` (framework-agnostic Bedrock SDK swap). A **bare**
|
||||
proprietary `openai.gpt-*` id (including bare `openai.gpt-5.5` / `openai.gpt-5.4`) is NEVER paired
|
||||
with a `converse`/`bedrock-runtime` path — bare ids are mantle ids. GPT-5.6, GPT-6, **GPT-5.5 and
|
||||
GPT-5.4** MAY use `runtime_openai_cris` on `bedrock-runtime` with a CRIS-prefixed id — this is
|
||||
not a violation (GPT-5.5 / GPT-5.4 runtime support: model cards read 2026-11-06, `us.`/`global.`
|
||||
ids only, no runtime in-region form). Note: `runtime_openai_cris` for GPT-6 Astra is valid
|
||||
only via Responses/Chat Completions/Converse — Astra does not support `Invoke`, so a design that
|
||||
needs Guardrails or Application Inference Profiles (Converse-only features that also require
|
||||
Invoke-adjacent plumbing on some other models) must verify against Astra's model card
|
||||
@@ -201,9 +204,10 @@ rubric file (`ai.md`, `ai-azure-openai-to-bedrock.md`, or a `vendored/ai/*` ref)
|
||||
`recommend_stay`, `honest_assessment_reason` names a non-cost blocker.
|
||||
- [ ] Bare proprietary GPT ids use Mantle; runtime Astra uses a supported `us.` / `global.`
|
||||
CRIS id and its caller-region gate. The recorded API and `model_change` match the design.
|
||||
- [ ] No `bedrock_models[]` entry pairs **GPT-5.5 or GPT-5.4** with a `converse`/`bedrock-runtime`
|
||||
migration path (those two models only — mantle-only, no CRIS). GPT-5.6 and GPT-6 models MAY
|
||||
use `migration_path: "runtime_openai_cris"` on `bedrock-runtime` — this is NOT a violation.
|
||||
- [ ] No `bedrock_models[]` entry pairs a **bare** proprietary GPT id (including bare
|
||||
`openai.gpt-5.5` / `openai.gpt-5.4`) with a `converse`/`bedrock-runtime` migration path.
|
||||
GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 MAY use `migration_path: "runtime_openai_cris"` on
|
||||
`bedrock-runtime` with a `us.`/`in.`/`global.` CRIS id — this is NOT a violation.
|
||||
- [ ] If `ai_source` is `azure_openai` or `openai`: every source model that is available on
|
||||
Bedrock and carried by the target region maps to ITSELF with `model_change: false` — not
|
||||
to a Claude/Nova substitute.
|
||||
|
||||
+4
-4
@@ -48,9 +48,9 @@ Inference on `bedrock-mantle` for these models is governed by **two per-model, p
|
||||
| Medium | Medium | Monitor 429s against **token** throughput, not request rate; enable prompt caching |
|
||||
| High | High | Enable prompt caching first (cached input is exempt from input TPM), then request a quota increase |
|
||||
|
||||
**The `bedrock-runtime` fallback exists only for GPT-5.6 and GPT-6 (Astra, Sol, Luna, 6.1 Sol, 6.1 Terra).** GPT-5.5 and GPT-5.4 are `bedrock-mantle` only and in-region only — for them, "switch to `bedrock-runtime`" requires moving to a different model (Bedrock-native or `gpt-oss`), a model change with its own eval cost. GPT-5.6 Sol/Terra/Luna and GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra DO have a `bedrock-runtime` path via CRIS inference profiles (GPT-5.6 `us.`/`in.`/`global.`, GPT-6 `us.`/`global.` prefixed ids; the model cards recommend runtime for new applications) — a legitimate endpoint option with its own quota family, and on Global CRIS it is also the cost-parity option. Scaling levers on the mantle path, in order:
|
||||
**Every proprietary GPT model on Bedrock has a `bedrock-runtime` path via CRIS inference profiles.** GPT-5.6 Sol/Terra/Luna (`us.`/`in.`/`global.`), GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra (`us.`/`global.`), and GPT-5.5 / GPT-5.4 (`us.`/`global.`; per their model cards read 2026-11-06, [GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) all list `bedrock-runtime` with Responses, Chat Completions and Converse; none has a runtime in-region form. The model cards recommend runtime for new applications. It is a legitimate endpoint option with its own quota family, and for GPT-5.6 / GPT-6 on Global CRIS it is also the cost-parity option (the GPT-5.5 / GPT-5.4 cards publish In-Region pricing only). A same-model endpoint switch is not a model change; only moving to a Bedrock-native model or `gpt-oss` carries the eval cost of a model change. Scaling levers on the mantle path, in order:
|
||||
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 only) — 90% off cached input and exempt from the input-TPM quota
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 on mantle; GPT-5.5 / GPT-5.4 list implicit caching on `bedrock-runtime` Responses only) — 90% off cached input and exempt from the input-TPM quota
|
||||
2. Exponential backoff with a bounded retry count (`max_retries` on the OpenAI SDK)
|
||||
3. Spreading load across minutes and ramping request rate gradually rather than bursting
|
||||
4. A Service Quotas increase for that model's input/output TPM in that region
|
||||
@@ -108,8 +108,8 @@ Claude models on Mantle have an additional **output TPM cap** that differs by mo
|
||||
|
||||
- For Claude migrations at medium/high volume: the 2M output TPM cap on Claude 4.7+ is the binding constraint
|
||||
- For OpenAI proprietary GPT targets: this Claude cap does not apply. Use the GPT-5.x mantle quotas or
|
||||
Astra-specific verification above. Astra / GPT-5.6 have runtime CRIS alternatives with their own
|
||||
quotas; GPT-5.5/5.4 have no runtime path
|
||||
Astra-specific verification above. Astra / GPT-5.6 / GPT-5.5 / GPT-5.4 all have runtime CRIS alternatives
|
||||
with their own quotas (`us.`/`global.` prefixed ids; no runtime in-region form)
|
||||
- For `gpt-oss` targets: these do run on `bedrock-runtime`, so standard account TPM limits and the Converse-path mitigations apply
|
||||
- When output-heavy workloads (long JSON, tool outputs, multi-step reasoning) are detected, flag the relevant cap prominently; recommend `bedrock-runtime` for production **only** when the target model actually has a `bedrock-runtime` path
|
||||
|
||||
|
||||
+32
-25
@@ -42,7 +42,9 @@ Two consequences that invert the previous decision logic:
|
||||
2. **The cost of a same-model move depends on the inference option — state it conditionally.** GPT-5.x In-Region and Geo
|
||||
CRIS are priced at OpenAI's _data residency_ tier, exactly 1.10x the standard list price, so those moves cost
|
||||
~10% more. **GPT-5.6 on Global CRIS is priced at the standard list price — cost parity** — available only when
|
||||
the workload has no data-residency constraint (GPT-5.5 / GPT-5.4 have no CRIS at all). See
|
||||
the workload has no data-residency constraint. GPT-5.5 / GPT-5.4 also have `us.`/`global.` CRIS ids on
|
||||
`bedrock-runtime`, but their model cards (read 2026-11-06) publish In-Region pricing only, so quote the ~10%
|
||||
figure for them on every inference option until a Global rate is published. See
|
||||
`shared/openai-on-bedrock.md`. The non-cost case — AWS commitments, IAM/VPC/CloudTrail governance, residency,
|
||||
prompt caching, one vendor relationship — carries the in-region path; never present that path as free or neutral.
|
||||
For Astra, use its separately dated Bedrock rates and verify the source provider's rate before comparing.
|
||||
@@ -72,14 +74,15 @@ GPT-5.4, the target is **the same model on Bedrock**.
|
||||
| GPT-5.6 Sol | GPT-5.6 Sol | `openai.gpt-5.6-sol` / `us.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Terra | GPT-5.6 Terra | `openai.gpt-5.6-terra` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Luna | GPT-5.6 Luna | `openai.gpt-5.6-luna` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
|
||||
Then apply the **region gate** below. Record `model_change: false` and the chosen path in
|
||||
`aws-design-ai.json` → `ai_architecture.code_migration`: `migration_path: "mantle_openai_chat"` for
|
||||
Astra sources using Chat Completions, or `"mantle_openai_responses"` for a selected Mantle Responses
|
||||
path. Use `migration_path: "runtime_openai_cris"` when a GPT-6 / GPT-5.6 target is served via
|
||||
`bedrock-runtime` with a CRIS id (the model cards recommend runtime for new applications).
|
||||
sources using Chat Completions on a model whose card lists it on mantle (GPT-6, GPT-5.5, GPT-5.4), or
|
||||
`"mantle_openai_responses"` for a selected Mantle Responses path. Use `migration_path: "runtime_openai_cris"`
|
||||
when a GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 target is served via `bedrock-runtime` with a CRIS id (the model
|
||||
cards recommend runtime for new applications).
|
||||
|
||||
Report the assessment honestly: `strong_migrate` here means "low-risk, well-supported move", not "cheaper".
|
||||
|
||||
@@ -105,14 +108,16 @@ regions — offer it, stating the data-residency implication of cross-region rou
|
||||
Global routes anywhere and is the cost-parity option). Only a workload that requires strictly in-region processing
|
||||
in a region outside the mantle matrix falls through to Tier 1.
|
||||
|
||||
**GPT-5.5 / GPT-5.4:** hard gate. Mantle in-region only, no CRIS, no fallback. If the target region is not in the
|
||||
matrix:
|
||||
**GPT-5.5 / GPT-5.4:** same shape as GPT-5.6, with a narrower mantle matrix (GPT-5.5: us-east-1, us-east-2;
|
||||
GPT-5.4: us-east-1, us-east-2, us-west-2, us-gov-west-1 — per their model cards, read 2026-11-06). If the target
|
||||
region is in that matrix, either endpoint works. Otherwise the `bedrock-runtime` CRIS path (`us.openai.gpt-5.5` /
|
||||
`global.openai.gpt-5.5`, `us.openai.gpt-5.4` / `global.openai.gpt-5.4`) reaches it, with the same data-residency
|
||||
caveat as GPT-5.6 (Geo stays within the geography, Global routes anywhere). There is no runtime in-region form.
|
||||
Only a workload that requires strictly in-region processing in a region outside the mantle matrix falls through:
|
||||
|
||||
1. Prefer offering a **region change** to a supported region, if no data-residency constraint pins the workload.
|
||||
This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. If the region is fixed, also offer **upgrading within the vendor to GPT-5.6** (which reaches the region via
|
||||
CRIS) alongside the cross-family option — a generation change, but same vendor and prompt idioms.
|
||||
3. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
1. Prefer offering a **region change** to a supported mantle region, if no data-residency constraint pins the
|
||||
workload. This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
|
||||
Record the outcome in `regional_warnings[]`.
|
||||
|
||||
@@ -235,10 +240,11 @@ model on price.
|
||||
|
||||
### Open-weight gpt-oss
|
||||
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and, unlike the
|
||||
proprietary GPT models, **do** support `bedrock-runtime` / Converse. They sit a capability class below the GPT-5.x
|
||||
frontier tier. Recommend them when the user wants OpenAI-architecture models on the Bedrock-native runtime surface
|
||||
(for Guardrails, invocation logging, or Converse-based tooling) rather than through mantle.
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and support
|
||||
`bedrock-runtime` / Converse with a bare model ID (no CRIS profile needed) and `Invoke`, which the proprietary GPT
|
||||
models do not. They sit a capability class below the GPT-5.x frontier tier. Recommend them when the user wants
|
||||
OpenAI-architecture models on the Bedrock-native runtime surface (for Guardrails, invocation logging, or
|
||||
Converse-based tooling) at the lowest price rather than through mantle.
|
||||
|
||||
---
|
||||
|
||||
@@ -284,7 +290,7 @@ rather than dismissing it. For a long-context (>272K) workload the gap is much w
|
||||
| Chat Completions | Preserve supported Chat when selected | Reshape only for a selected API change; verify the model and endpoint |
|
||||
| Function calling | Supported on GPT-5.x via mantle; Claude tools via Converse | Same-model keeps tool semantics identical |
|
||||
| Reasoning effort | `reasoning={"effort": ...}` on mantle | `none` / `low` / `medium` / `high` / `xhigh` / `max` |
|
||||
| Prompt caching | GPT-5.6 only (90% off cached input) | Not listed for GPT-5.5 / 5.4; Claude has its own caching |
|
||||
| Prompt caching | GPT-5.6 (90% off cached input); GPT-5.5 / 5.4 implicit caching on `bedrock-runtime` Responses only | GPT-5.5 / 5.4 cards (read 2026-11-06) list a cache-read rate and implicit caching on runtime, none on mantle; Claude has its own caching |
|
||||
| Streaming | Supported | Verify per surface |
|
||||
| Vision | GPT-5.x (image input) or Claude / Llama 4 | Same-model preserves behavior |
|
||||
| JSON mode | Claude (excellent), Nova Pro (good) | Most models via prompt |
|
||||
@@ -339,11 +345,12 @@ and closed to new customers as of July 30, 2026.
|
||||
### When to prefer a Bedrock-native model over the mantle GPT path
|
||||
|
||||
Choose Claude / Nova via Converse when the workload needs Knowledge Bases, intelligent prompt routing, structured
|
||||
outputs, or application inference profiles — none of which the GPT path offers. Note what no longer forces a family
|
||||
switch for GPT-5.6: its `bedrock-runtime` CRIS path provides cross-region inference, invocation logs, and
|
||||
Guardrails (Converse API only; prompt caching is Responses-only on runtime). For GPT-5.5 / GPT-5.4, Guardrails,
|
||||
logging, and cross-region reach still require a different model. The mantle GPT path is the smallest change; the
|
||||
Bedrock-native path remains the most feature-complete.
|
||||
outputs, or application inference profiles — none of which the mantle GPT path offers. Note what no longer forces
|
||||
a family switch for GPT-5.6, GPT-5.5 and GPT-5.4: their `bedrock-runtime` CRIS paths provide cross-region
|
||||
inference, invocation logs, Guardrails and application inference profiles (both Converse API only), and
|
||||
structured outputs and implicit prompt caching (both Responses API only) — per the GPT-5.4 / GPT-5.5 model cards
|
||||
read 2026-11-06. Knowledge Bases and intelligent prompt routing still require a Bedrock-native model. The mantle
|
||||
GPT path is the smallest change; the Bedrock-native path remains the most feature-complete.
|
||||
|
||||
### High-spend tiering
|
||||
|
||||
@@ -364,8 +371,8 @@ alternative, not against the source provider.
|
||||
cheaper and exempt from the input-TPM quota, which relieves the binding constraint on mantle.
|
||||
|
||||
**Very high (>100M tokens/day):** tiering plus caching. Check TPM headroom per model per region early; there is no
|
||||
RPM quota and no `bedrock-runtime` fallback for GPT models, so quota increases are the scaling lever. See
|
||||
`shared/ai-migration-guardrails.md`.
|
||||
RPM quota, and the `bedrock-runtime` CRIS path for GPT-5.6 / 5.5 / 5.4 / GPT-6 carries its own quotas rather than
|
||||
adding headroom to mantle, so quota increases are the scaling lever. See `shared/ai-migration-guardrails.md`.
|
||||
|
||||
---
|
||||
|
||||
|
||||
+20
-12
@@ -133,8 +133,8 @@ See `vendored/ai/ai-model-lifecycle.md` for lifecycle details. **Do not recommen
|
||||
| GPT-5.6 Sol | openai.gpt-5.6-sol | OpenAI | 4.40 | 22.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 8.80/33.00) |
|
||||
| GPT-5.6 Terra | openai.gpt-5.6-terra | OpenAI | 2.20 | 13.20 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 4.40/19.80) |
|
||||
| GPT-5.6 Luna | openai.gpt-5.6-luna | OpenAI | 0.22 | 1.32 | 272K | fast | active (mantle in-region + runtime CRIS; 1M tier 0.44/1.98) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle only; no 1M tier) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle only; no 1M tier) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 11.00/49.50; card read 2026-11-06) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 5.50/24.75; card read 2026-11-06) |
|
||||
| Gemma 4 31B | google.gemma-4-31b | Google | 0.14 | 0.40 | 256K | mid | active (Mantle only) |
|
||||
| Gemma 4 26B-A4B | google.gemma-4-26b-a4b | Google | 0.13 | 0.40 | 256K | efficient | active (Mantle only) |
|
||||
| Gemma 4 E2B | google.gemma-4-e2b | Google | 0.04 | 0.08 | 128K | budget | active (Mantle only) |
|
||||
@@ -353,17 +353,24 @@ Regions (mantle): us-east-1 / us-east-2 for all five; us-west-2 additionally car
|
||||
| GPT-5.5 | 5.50 | 33.00 | — | 0.55 |
|
||||
| GPT-5.4 | 2.75 | 16.50 | — | 0.275 |
|
||||
|
||||
#### Long context (1M) — GPT-5.6 only, 2.0x input / 1.5x output
|
||||
#### Long context (>272K input tokens) — GPT-5.6 at 2.0x input / 1.5x output; GPT-5.5 / GPT-5.4 per their model cards
|
||||
|
||||
| Model | Input $/1M | Output $/1M |
|
||||
| ------------- | ---------- | ----------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 |
|
||||
| Model | Input $/1M | Output $/1M | Cache read $/1M |
|
||||
| ------------- | ---------- | ----------- | --------------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 | — |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 | — |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 | — |
|
||||
| GPT-5.5 | 11.00 | 49.50 | 1.10 |
|
||||
| GPT-5.4 | 5.50 | 24.75 | 0.55 |
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier. GPT-5.5 and GPT-5.4 have no such tier, so
|
||||
their usable window is 272K. Global CRIS is published at OpenAI's standard list price (cost parity, GPT-5.6 only —
|
||||
see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
A workload above 272K input tokens must be priced at the long-context tier, applied to all input and output tokens
|
||||
of the request. GPT-5.5 and GPT-5.4 rows read from their model cards on 2026-11-06
|
||||
([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html),
|
||||
[GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)); both cards state a
|
||||
1,050,000-token context window (128,000 max output) and list `us.`/`global.` CRIS ids on `bedrock-runtime`, but
|
||||
publish the In-Region inference option only. Global CRIS is published at OpenAI's standard list price (cost parity,
|
||||
GPT-5.6 only — see `openai-on-bedrock.md`). GovCloud (US-West) GPT-5.4 short-context is 3.375 / 20.25 (cache read
|
||||
0.3375) per the 2026-11-06 card; an earlier revision recorded 3.30 / 19.80.
|
||||
|
||||
> **One source conflict.** These models are absent
|
||||
> from the AWS Price List API entirely: querying `AmazonBedrock` returns only `gpt-oss` and GPT OSS Safeguard, and
|
||||
@@ -373,7 +380,8 @@ see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
|
||||
**Prompt caching (GPT-5.6 only):** cached input read at a 90% discount, cache write at 1.25x uncached input, minimum
|
||||
1,024-token prefix, up to 4 breakpoints, ≥30-minute retention. Cached input is exempt from the input-TPM quota.
|
||||
GPT-5.5 and GPT-5.4 do not list prompt caching support.
|
||||
GPT-5.5 and GPT-5.4 list only implicit prompt caching on `bedrock-runtime` Responses (automatic, no request
|
||||
parameters; cache-read rate above, no cache-write rate) and none on mantle (model cards read 2026-11-06).
|
||||
|
||||
#### Open-weight gpt-oss
|
||||
|
||||
|
||||
+15
-8
@@ -190,12 +190,18 @@ Traditional-AI capabilities (AWS AI service, not Bedrock): `document_extraction`
|
||||
## migration_path vocabulary
|
||||
|
||||
`mantle_openai_responses` (an OpenAI-protocol source — direct or Azure-fronted — with the model
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `converse`
|
||||
(Bedrock-native Converse/`bedrock-runtime`; set `model_change: true` when a proprietary GPT
|
||||
model must move off mantle for Guardrails / Knowledge Bases / logging / an unsupported region) ·
|
||||
`gpt-oss` (OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic
|
||||
Bedrock SDK swap). A proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — those models are mantle-only.
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `mantle_openai_chat`
|
||||
(same, preserving a Chat Completions source on a model whose card lists Chat Completions on
|
||||
mantle — GPT-6, GPT-5.5, GPT-5.4) · `runtime_openai_cris` (same-model move served via
|
||||
`bedrock-runtime` with a `us.`/`in.`/`global.` CRIS id, `model_change: false`; available for
|
||||
GPT-6, GPT-5.6, GPT-5.5 and GPT-5.4 per their model cards) · `converse` (Bedrock-native
|
||||
Converse/`bedrock-runtime`; set `model_change: true` when the workload needs Knowledge Bases,
|
||||
intelligent prompt routing, or another feature no same-model path serves) · `gpt-oss`
|
||||
(OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic Bedrock SDK
|
||||
swap). A **bare** proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — bare ids are mantle ids; every proprietary GPT target on
|
||||
`bedrock-runtime` carries a CRIS-prefixed id (there is no runtime in-region form for GPT-6,
|
||||
GPT-5.6, GPT-5.5 or GPT-5.4).
|
||||
|
||||
## regional_validation
|
||||
|
||||
@@ -255,8 +261,9 @@ rubric file (`ai.md`, a provider-to-Bedrock mapping guide, or a `vendored/ai/*`
|
||||
`price_comparison`.
|
||||
- [ ] `honest_assessment` overall is the weakest across `bedrock_models[]`; if
|
||||
`recommend_stay`, `honest_assessment_reason` names a non-cost blocker.
|
||||
- [ ] No `bedrock_models[]` entry pairs a proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (mantle-only).
|
||||
- [ ] No `bedrock_models[]` entry pairs a **bare** proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (bare ids are mantle ids; a `bedrock-runtime`
|
||||
target for GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 carries a `us.`/`in.`/`global.` CRIS id).
|
||||
- [ ] For a retained source whose `source_service` is `azure_openai`, or whose provider resolves
|
||||
to direct OpenAI: every source model that is available on Bedrock and carried by the
|
||||
target region maps to ITSELF with `model_change: false` — not to a Claude/Nova substitute.
|
||||
|
||||
+2
-2
@@ -225,8 +225,8 @@ _Skip when:_ the validator summary reports `models_extractable: true` — every
|
||||
|
||||
**Same-model rows first.** GPT-6 Astra / GPT-5.6 / 5.5 / 5.4 sources map to themselves.
|
||||
Design applies the model-specific API and region gate; Clarify records the baseline only.
|
||||
Astra Standard mantle supports `us-east-1` and `us-west-2`; runtime uses US Geo / Global CRIS. GPT-5.6 also has runtime CRIS;
|
||||
GPT-5.5 / 5.4 remain mantle-only. GPT-5.x in-region costs about 10% above OpenAI standard,
|
||||
Astra Standard mantle supports `us-east-1` and `us-west-2`; runtime uses US Geo / Global CRIS. GPT-5.6, GPT-5.5
|
||||
and GPT-5.4 also have runtime CRIS (`us.` / `global.`). GPT-5.x in-region costs about 10% above OpenAI standard,
|
||||
while GPT-5.6 Global CRIS uses the standard rate. For Astra, compare its separately dated
|
||||
inference/context tier against the source price before claiming savings or parity.
|
||||
For sources without a Bedrock equivalent, present a same-vendor upgrade and a cross-family option.
|
||||
|
||||
+2
-2
@@ -204,8 +204,8 @@ _Skip when:_ `models[].model_id` is populated in `ai-workload-profile.json` **wi
|
||||
|
||||
**Same-model rows first.** GPT-6 Astra / GPT-5.6 / 5.5 / 5.4 sources map to themselves.
|
||||
Apply the model-specific API and region gate in `references/shared/openai-on-bedrock.md`.
|
||||
Astra Standard mantle supports `us-east-1` and `us-west-2`; runtime uses US Geo / Global CRIS. GPT-5.6 also has runtime CRIS;
|
||||
GPT-5.5 / 5.4 remain mantle-only. GPT-5.x in-region costs about 10% above OpenAI standard,
|
||||
Astra Standard mantle supports `us-east-1` and `us-west-2`; runtime uses US Geo / Global CRIS. GPT-5.6, GPT-5.5
|
||||
and GPT-5.4 also have runtime CRIS (`us.` / `global.`). GPT-5.x in-region costs about 10% above OpenAI standard,
|
||||
while GPT-5.6 Global CRIS uses the standard rate. For Astra, compare its separately dated
|
||||
inference/context tier against the source price before claiming savings or parity.
|
||||
For sources without a Bedrock equivalent, present a same-vendor upgrade and a cross-family option.
|
||||
|
||||
+20
-15
@@ -238,7 +238,9 @@ data-residency-tier premium, while supported Global CRIS uses standard-rate pari
|
||||
Report the actual option; do not downgrade solely for the modest verified premium.
|
||||
**GPT-5.6/GPT-6 only:** above 272K context, apply 2.0x input and 1.5x output rates only when
|
||||
the selected model and inference option support that tier; the resulting gap may change
|
||||
the assessment. GPT-5.5/5.4 have no 1M tier. **Opus 5.5:** use the selected region/profile's
|
||||
the assessment. GPT-5.5/5.4 have their own long-context (>272K) rates on their model cards
|
||||
(read 2026-11-06; see `openai-on-bedrock.md` § Pricing) — use those, not the GPT-5.6
|
||||
multiplier. **Opus 5.5:** use the selected region/profile's
|
||||
flat rates from `references/vendored/ai/bedrock-pricing-cache.md` throughout the full 1M context;
|
||||
do not apply the GPT multiplier. Verify other models' tiers individually.
|
||||
|
||||
@@ -407,23 +409,25 @@ Astra Chat solely because it is proprietary GPT. Record `model_change: false` on
|
||||
the exact source model is retained; a Pro-to-Astra upgrade is a model change requiring eval.
|
||||
GPT-5.x retains its separately dated API evidence and any selected Chat-to-Responses reshape.
|
||||
|
||||
**A Chat Completions source is a supported near-drop-in for GPT-6 and GPT-5.6 — not a mandatory reshape.** All five GPT-6 models and the GPT-5.6 family list Chat Completions as a supported API on both mantle and runtime CRIS (per `openai-on-bedrock.md` § Access Paths); if `integration` evidence shows `chat.completions.create` for one of these models, the design may keep `chat.completions.create`/`choices[0].message` parsing — reshape to `responses.create` is optional, only for a Responses-only feature (runtime explicit prompt caching, reasoning-item round-trips) or a demonstrated constraint. **GPT-5.5 and GPT-5.4 are the exception**: Chat Completions support is unconfirmed on their model cards (Responses on mantle is the only verified surface), so a Chat Completions source on either model still needs a planned reshape, probed against the target account before committing.
|
||||
**A Chat Completions source is a supported near-drop-in for GPT-6 and GPT-5.6 — not a mandatory reshape.** All five GPT-6 models and the GPT-5.6 family list Chat Completions as a supported API on both mantle and runtime CRIS (per `openai-on-bedrock.md` § Access Paths); if `integration` evidence shows `chat.completions.create` for one of these models, the design may keep `chat.completions.create`/`choices[0].message` parsing — reshape to `responses.create` is optional, only for a Responses-only feature (runtime explicit prompt caching, reasoning-item round-trips) or a demonstrated constraint. **GPT-5.5 and GPT-5.4 follow the same rule**: their model cards (read 2026-11-06) list Chat Completions as supported on both `bedrock-mantle` and `bedrock-runtime`, so a Chat Completions source on either model may keep Chat Completions; the reshape is optional for them too. A planned reshape is required only when the selected model's card leaves Chat Completions unchecked on the selected endpoint (check the per-model path entries in the canonical catalog). Probe the selected API against the target account before committing, as for any model.
|
||||
|
||||
**Terra caching depends on the endpoint and mode.** On runtime, implicit caching supports
|
||||
Responses, Chat Completions, InvokeModel, and Converse. It does not require a Chat client
|
||||
reshape and can be combined with Converse Guardrails. Runtime explicit caching requires
|
||||
Responses. On Mantle, both implicit and explicit caching are Responses-only.
|
||||
|
||||
**No Converse fallback for GPT-5.5/GPT-5.4 — GPT-5.6 and GPT-6 differ.** GPT-5.5 and GPT-5.4 are
|
||||
`bedrock-mantle` only and in-region only — there is no `bedrock-runtime` path and no cross-region
|
||||
inference profile for these two models. GPT-5.6 Sol/Terra/Luna and all five GPT-6 models (Astra,
|
||||
Sol, Luna, 6.1 Sol, 6.1 Terra) DO have a `bedrock-runtime` CRIS path (`us.`/`global.` prefixed IDs, per
|
||||
`openai-on-bedrock.md` § Access Paths) — for these models, a region or feature need (Guardrails,
|
||||
Knowledge Bases, invocation logging, or a region outside the mantle matrix) is satisfied by
|
||||
**switching endpoint, not model**: record `migration_path: "runtime_openai_cris"` with
|
||||
`model_change: false`. Reserve `migration_path: "converse"` with `model_change: true` for
|
||||
GPT-5.5/GPT-5.4 only, or when the GPT-6/5.6 CRIS path itself cannot serve the need (e.g. a feature
|
||||
CRIS does not support — see the runtime-side feature splits in `openai-on-bedrock.md`). Note:
|
||||
**Every proprietary GPT model on Bedrock has a `bedrock-runtime` CRIS path — including GPT-5.5 and
|
||||
GPT-5.4.** GPT-5.6 Sol/Terra/Luna, all five GPT-6 models (Astra, Sol, Luna, 6.1 Sol, 6.1 Terra), and
|
||||
GPT-5.5 / GPT-5.4 (model cards read 2026-11-06: `us.openai.gpt-5.5` / `global.openai.gpt-5.5`,
|
||||
`us.openai.gpt-5.4` / `global.openai.gpt-5.4`; Responses, Chat Completions and Converse on runtime)
|
||||
have a `bedrock-runtime` CRIS path (`us.`/`global.` prefixed IDs, per `openai-on-bedrock.md`
|
||||
§ Access Paths). No proprietary GPT model has a runtime in-region form. For all of them, a region
|
||||
or feature need (Guardrails, invocation logging, application inference profiles, or a region
|
||||
outside the mantle matrix) is satisfied by **switching endpoint, not model**: record
|
||||
`migration_path: "runtime_openai_cris"` with `model_change: false`. Reserve
|
||||
`migration_path: "converse"` with `model_change: true` for a need the same-model CRIS path itself
|
||||
cannot serve (Knowledge Bases, intelligent prompt routing, or a feature CRIS does not support —
|
||||
see the runtime-side feature splits in `openai-on-bedrock.md`). Note:
|
||||
`runtime_openai_cris` for GPT-6 Astra is valid only via Responses/Chat Completions/Converse —
|
||||
Astra does not support `Invoke`, so a design that needs Guardrails or Application Inference
|
||||
Profiles (Converse-only features that also require Invoke-adjacent plumbing on some other models)
|
||||
@@ -511,9 +515,10 @@ Write `aws-design-ai.json` to `$MIGRATION_DIR/`.
|
||||
- [ ] `honest_assessment` logic is consistent (weakest model drives overall)
|
||||
- [ ] If `ai_source` includes `"openai"`: every model that `shared/openai-on-bedrock.md` lists as available on Bedrock, and whose target region carries it, is mapped to **itself** with `model_change: false` — not to a Claude/Nova substitute
|
||||
- [ ] If `honest_assessment == "recommend_stay"`: `honest_assessment_reason` names a non-cost blocker (cost parity alone is never sufficient for an OpenAI source)
|
||||
- [ ] No `bedrock_models[]` entry pairs **GPT-5.5 or GPT-5.4** with a Converse/`bedrock-runtime`
|
||||
migration path (those two models only — mantle-only, no CRIS). GPT-5.6 and GPT-6 models MAY
|
||||
use `migration_path: "runtime_openai_cris"` on `bedrock-runtime` — this is NOT a violation.
|
||||
- [ ] No `bedrock_models[]` entry pairs a **bare** proprietary GPT id (including bare
|
||||
`openai.gpt-5.5` / `openai.gpt-5.4`) with a Converse/`bedrock-runtime` migration path.
|
||||
GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 MAY use `migration_path: "runtime_openai_cris"` on
|
||||
`bedrock-runtime` with a `us.`/`in.`/`global.` CRIS id — this is NOT a violation.
|
||||
- [ ] `regional_warnings` is present (empty array `[]` if no issues; populated if any service unavailable in target region)
|
||||
- [ ] `multi_model_warnings` is present (empty array `[]` if single model or no coordination issues)
|
||||
- [ ] If `agentic_profile.is_agentic == true`: `agentic_design` object is present with `migration_approach` matching `preferences.json`
|
||||
|
||||
+2
-2
@@ -104,8 +104,8 @@ See `references/shared/openai-on-bedrock.md` for the model-specific pricing rule
|
||||
|
||||
For the GPT-5.x rows, the mantle endpoint is in-region only (us-east-1, us-east-2; us-west-2 additionally for
|
||||
Terra, Luna, and GPT-5.4; AWS GovCloud us-gov-west-1 / us-gov-east-1 for Terra and Luna, us-gov-west-1 also for
|
||||
GPT-5.4). GPT-5.6 additionally reaches most commercial regions via `bedrock-runtime` CRIS ids; GPT-5.5 / GPT-5.4
|
||||
have no CRIS.
|
||||
GPT-5.4). GPT-5.6, GPT-5.5 and GPT-5.4 additionally reach most commercial regions via `bedrock-runtime` CRIS ids
|
||||
(`us.` / `global.`; GPT-5.5 / GPT-5.4 per their model cards read 2026-11-06).
|
||||
|
||||
For the GPT-6 rows, mantle is us-east-1 only, except GPT-6 Astra which is also available in us-west-2. All five
|
||||
GPT-6 models additionally reach a broad set of commercial regions via `bedrock-runtime` CRIS (`us.` / `global.`
|
||||
|
||||
+2
-2
@@ -32,7 +32,7 @@ Check `preferences.json` → `ai_constraints.ai_framework` to determine the migr
|
||||
|
||||
- `mantle` / `mantle_openai_responses`: point the client at `https://bedrock-mantle.{region}.api.aws/openai/v1` — note the `openai/v1` segment for proprietary GPT models; a bare `/v1` 404s. Swap the credential for a Bedrock API key or the auto-refreshing token provider — **not** the existing OpenAI key. Update the model string to the bare Bedrock model ID (`openai.gpt-*` for a same-model move). Grant the `bedrock-mantle:*` actions; `bedrock:InvokeModel` does not authorize these models.
|
||||
- `runtime_openai_cris` (GPT-5.6/GPT-6 via `bedrock-runtime` CRIS): point the client at `https://bedrock-runtime.{region}.amazonaws.com/openai/v1` instead — **not** the mantle URL. Update the model string to the `us.`/`in.`/`global.`-prefixed CRIS inference-profile ID, not the bare ID. Grant standard `bedrock:InvokeModel*` actions against the CRIS inference-profile ARN — **not** `bedrock-mantle:*`. For a GPT-6 CRIS id, also grant `bedrock:InvokeModel` on `arn:aws:bedrock:{region}:{account}:project/default` alongside the inference-profile ARN (per the GPT-6 model cards). Check `aws-design-ai.json` → `ai_architecture.bedrock_models[].required_client_api`: if non-Responses, emit that surface (e.g. Converse) instead of defaulting to `responses.create` — absent/`responses` keeps the Responses surface for a region-only move.
|
||||
- **Config-only if the source already calls `responses.create`.** A Chat Completions source for GPT-6 or GPT-5.6 may keep `chat.completions.create`/`choices[0].message` parsing unchanged — Chat Completions is supported on both mantle and runtime CRIS for these models, so reshape to Responses is optional (only for a Responses-only feature or demonstrated constraint). For GPT-5.5/GPT-5.4, Chat Completions support is unconfirmed, so still plan a reshape to Responses and probe the target account before committing.
|
||||
- **Config-only if the source already calls `responses.create`.** A Chat Completions source for GPT-6 or GPT-5.6 may keep `chat.completions.create`/`choices[0].message` parsing unchanged — Chat Completions is supported on both mantle and runtime CRIS for these models, so reshape to Responses is optional (only for a Responses-only feature or demonstrated constraint). GPT-5.5/GPT-5.4 follow the same rule — their model cards (read 2026-11-06) list Chat Completions on both mantle and runtime, so the reshape is optional for them too; probe the selected API against the target account before committing, as for any model.
|
||||
- Test in staging, validate responses
|
||||
|
||||
Match the `migration_path` prefix, not the exact string: Design writes the more specific `mantle_openai_responses` for a same-model OpenAI migration, and an equality test against `mantle` would misroute those runs into the adapter path below. `runtime_openai_cris` (GPT-5.6/GPT-6 served via `bedrock-runtime` CRIS instead of `bedrock-mantle`) is matched explicitly alongside the `mantle*` prefix — it does not start with `mantle`, so it needs its own check to avoid falling through to the adapter path below.
|
||||
@@ -55,7 +55,7 @@ Based on `ai-workload-profile.json` → `integration.pattern` and `integration.l
|
||||
|
||||
| Source SDK | Target | Key Change |
|
||||
| ----------------------- | --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
|
||||
| OpenAI SDK (same model) | Mantle Responses API | Base URL → `.../openai/v1`, Bedrock token/API key (not an OpenAI key), `openai.gpt-*` model ID. Config-only **if** already on `responses.create`; a Chat Completions source may keep Chat Completions for GPT-6/GPT-5.6 (reshape to Responses optional); GPT-5.5/GPT-5.4 still need a reshape. |
|
||||
| OpenAI SDK (same model) | Mantle Responses API | Base URL → `.../openai/v1`, Bedrock token/API key (not an OpenAI key), `openai.gpt-*` model ID. Config-only **if** already on `responses.create`; a Chat Completions source may keep Chat Completions for GPT-6/GPT-5.6/GPT-5.5/GPT-5.4 (reshape to Responses optional; plan it only when the selected model's card leaves Chat Completions unchecked on the selected endpoint). |
|
||||
| Vertex AI (Python) | boto3 Bedrock Converse API | `GenerativeModel.generate_content()` → `bedrock.converse()` |
|
||||
| Vertex AI (JS) | @aws-sdk/client-bedrock-runtime | `model.generateContent()` → `client.send(new ConverseCommand())` |
|
||||
| Vertex AI (Go) | aws-sdk-go-v2 bedrockruntime | `aiplatform` → `bedrockruntime.Converse()` |
|
||||
|
||||
+62
-29
@@ -2,6 +2,7 @@
|
||||
|
||||
**Last verified:** 2026-10-10
|
||||
**GPT-6.1 Terra verified:** 2026-10-27
|
||||
**GPT-5.5 / GPT-5.4 re-verified:** 2026-11-06 against their model cards ([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) — both now list `bedrock-runtime` (Responses, Chat Completions, Converse via `us.`/`global.` CRIS ids), Chat Completions on mantle, a 1.05M context window and a long-context price tier
|
||||
**Sources:** [OpenAI model cards](https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html) (per-model
|
||||
cards linked below), [GPT-5.6 launch post](https://aws.amazon.com/blogs/machine-learning/get-started-with-openai-gpt-5-6-sol-terra-and-luna-on-amazon-bedrock/),
|
||||
[GPT-5.6 GA announcement](https://aws.amazon.com/about-aws/whats-new/2026/07/openai-gpt-sol-terra/),
|
||||
@@ -32,8 +33,8 @@ risk) all defer to it. Do not restate model IDs, regions, or endpoint paths else
|
||||
| GPT-5.6 Sol | `openai.gpt-5.6-sol` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html) |
|
||||
| GPT-5.6 Terra | `openai.gpt-5.6-terra` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-terra.html) |
|
||||
| GPT-5.6 Luna | `openai.gpt-5.6-luna` | Jul 13, 2026 | 1M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-luna.html) |
|
||||
| GPT-5.5 | `openai.gpt-5.5` | Jun 1, 2026 | 272K | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html) |
|
||||
| GPT-5.4 | `openai.gpt-5.4` | Jun 1, 2026 | 272K | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html) |
|
||||
| GPT-5.5 | `openai.gpt-5.5` | Jun 1, 2026 | 1.05M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html) |
|
||||
| GPT-5.4 | `openai.gpt-5.4` | Jun 1, 2026 | 1.05M | Active | [card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html) |
|
||||
| gpt-oss-120b | `openai.gpt-oss-120b` | Aug 5, 2025 | 128K | Active | open-weight; also on `bedrock-runtime` as `openai.gpt-oss-120b-1:0` |
|
||||
| gpt-oss-20b | `openai.gpt-oss-20b` | Aug 5, 2025 | 128K | Active | open-weight; also on `bedrock-runtime` as `openai.gpt-oss-20b-1:0` |
|
||||
| GPT OSS Safeguard | `openai.gpt-oss-safeguard-120b` / `-20b` | — | — | Active | content-moderation / guardrail enforcement, not general chat |
|
||||
@@ -50,8 +51,10 @@ GPT-6 Astra is the newest, most capable OpenAI model in this catalog. Sol remain
|
||||
lower-cost reasoning alternative; Astra does not make existing GPT-5.x sources require a model upgrade.
|
||||
|
||||
> **Context-window conflict (resolved):** the GPT-5.6 launch blog states 272K for all three variants; all three
|
||||
> model cards state 1M. **The model cards are authoritative** — use 1M for GPT-5.6. GPT-5.5 and GPT-5.4 are 272K on
|
||||
> both sources. Re-check on refresh; if AWS corrects the blog, the cards still win.
|
||||
> model cards state 1M. **The model cards are authoritative** — use 1M for GPT-5.6. The GPT-5.5 and GPT-5.4 cards
|
||||
> (read 2026-11-06) state 1,050,000 tokens shared between input and output with 128,000 max output; the 272K figure
|
||||
> recorded earlier is now only their short-/long-context pricing breakpoint. Re-check on refresh; if AWS corrects
|
||||
> the blog, the cards still win.
|
||||
|
||||
**Not on Bedrock (as of this refresh):** GPT-4o, GPT-4.1, GPT-4 / GPT-4 Turbo, GPT-3.5 Turbo, the o-series
|
||||
(o1/o3/o4-mini), GPT-5 / GPT-5.1 / GPT-5.2, Codex (unverified — see the Codex note below), and the `*-Pro` variants
|
||||
@@ -116,8 +119,12 @@ input and text output. The launch announcement rounds the input context to 1M; u
|
||||
exact limits for capacity checks. Browser/computer-use capability does not establish parity with
|
||||
OpenAI-hosted tools; probe the application's specific tool flow.
|
||||
|
||||
**GPT-5.5 and GPT-5.4 are `bedrock-mantle`-only and in-region only.** Their model cards list a single
|
||||
Programmatic Access row (`bedrock-mantle`, Geo/Global "Not supported") and In-Region pricing only.
|
||||
**GPT-5.5 and GPT-5.4 have TWO endpoints** (re-verified 2026-11-06 against their model cards; earlier revisions of
|
||||
this file recorded them as `bedrock-mantle`-only). Both Programmatic Access tables now carry a `bedrock-runtime`
|
||||
row (In-Region "Not supported"; Geo `us.openai.gpt-5.5` / `us.openai.gpt-5.4`; Global `global.openai.gpt-5.5` /
|
||||
`global.openai.gpt-5.4`) and a `bedrock-mantle` row (bare id, in-region only). Supported APIs: `Responses`,
|
||||
`Chat Completions` and `Converse` on `bedrock-runtime`; `Responses` and `Chat Completions` on `bedrock-mantle`;
|
||||
`Messages` and `Invoke` on neither. The pricing tables still publish the In-Region inference option only.
|
||||
|
||||
**GPT-5.6 Sol / Terra / Luna have TWO endpoints** (verified 2026-08-21 — this changed after this file's original
|
||||
2026-08-10 verification; the Refresh Checklist predicted it):
|
||||
@@ -136,13 +143,16 @@ Constraints that still break naive assumptions:
|
||||
mantle path explicitly, and the runtime example gives `bedrock-runtime.{region}.amazonaws.com/openai/v1`.
|
||||
2. **The endpoints take different model-id forms.** Mantle takes the bare `openai.gpt-5.6-*` id and has no CRIS;
|
||||
`bedrock-runtime` takes ONLY a CRIS inference-profile id (`us.` / `in.` / `global.` prefixed) and has no
|
||||
in-region form. `bedrock:ListInferenceProfiles` returns the 5.6 CRIS profiles; it never returns the bare mantle
|
||||
ids, and it returns nothing for GPT-5.5 / GPT-5.4.
|
||||
in-region form. `bedrock:ListInferenceProfiles` returns the CRIS profiles (5.6, and `us.`/`global.` for
|
||||
GPT-5.5 / GPT-5.4 per their cards); it never returns the bare mantle ids.
|
||||
3. **API surfaces differ by endpoint.** The 5.6 cards list `Responses`, `Chat Completions`, `Invoke`, and
|
||||
`Converse` as supported APIs, with runtime-side feature splits: **Guardrails are Converse-only; prompt caching is
|
||||
Responses-only on runtime**; server-side tool use, structured outputs, and application inference profiles are NOT
|
||||
supported on runtime (server-side tool calling IS supported on mantle). For GPT-5.5 / GPT-5.4 treat Responses on
|
||||
mantle as the only verified surface.
|
||||
supported on runtime (server-side tool calling IS supported on mantle). The GPT-5.5 / GPT-5.4 cards (read
|
||||
2026-11-06) list `Responses`, `Chat Completions` and `Converse` on runtime (no `Invoke`) and `Responses` and
|
||||
`Chat Completions` on mantle; on runtime they list Guardrails and application inference profiles as
|
||||
Converse-only, and implicit prompt caching and structured outputs as Responses-only. Chat Completions is a
|
||||
documented surface for GPT-5.5 / GPT-5.4 on both endpoints — probe it in the target account like any other.
|
||||
|
||||
### GPT-6 family: Ultrafast tier
|
||||
|
||||
@@ -193,8 +203,10 @@ that expires within 12 hours and is not refreshed. Do not recommend it for produ
|
||||
|
||||
**IAM:** on the mantle path, the managed policy `AmazonBedrockMantleInferenceAccess` grants what inference needs,
|
||||
including `bedrock-mantle:CreateInference` and `bedrock-mantle:CallWithBearerToken` — `bedrock:InvokeModel` does not
|
||||
authorize mantle calls. On the GPT-5.6 `bedrock-runtime` path the usual `bedrock:InvokeModel*` against the CRIS
|
||||
inference-profile ARN applies, as for any other runtime model.
|
||||
authorize mantle calls. On the GPT-5.6 / GPT-5.5 / GPT-5.4 `bedrock-runtime` path the usual `bedrock:InvokeModel*`
|
||||
against the CRIS inference-profile ARN applies, as for any other runtime model; the GPT-5.5 / GPT-5.4 cards add
|
||||
that the identity also needs `bedrock:InvokeModel` on the account's default project
|
||||
(`arn:aws:bedrock:{region}:{account-id}:project/default`).
|
||||
|
||||
**Reasoning effort:** the five GPT-5.x models accept `none`, `low`, `medium`, `high`, `xhigh`, `max`. Astra's card
|
||||
does not enumerate these values; verify its accepted settings rather than inheriting the GPT-5.x list. Because these models reason
|
||||
@@ -205,7 +217,8 @@ request for multi-turn and tool-calling flows.
|
||||
|
||||
## Regional Availability — Endpoint-Aware
|
||||
|
||||
**The mantle in-region matrix** (the only reach for GPT-5.5 / GPT-5.4, and the in-region option for GPT-6 Astra / GPT-5.6):
|
||||
**The mantle in-region matrix** (the in-region option for every proprietary GPT model; GPT-5.5 / GPT-5.4 rows
|
||||
re-verified 2026-11-06):
|
||||
|
||||
| Model | us-east-1 | us-east-2 | us-west-2 | us-gov-west-1 | us-gov-east-1 |
|
||||
| ------------- | --------- | --------- | --------- | ------------- | ------------- |
|
||||
@@ -235,8 +248,16 @@ any GPT-6 model, it just means using the runtime endpoint with a CRIS id.
|
||||
**GPT-5.6 additionally reaches most commercial regions via `bedrock-runtime` CRIS** (Geo `us.` / `in.`, Global
|
||||
`global.` inference profiles; the Sol card's runtime footprint spans 30+ regions). So a region outside the mantle
|
||||
matrix does NOT make the same-model path unavailable for GPT-5.6 — it means using the runtime endpoint with a CRIS
|
||||
id, with the data-residency implications of cross-region routing. For GPT-5.5 / GPT-5.4 the mantle matrix is a hard
|
||||
gate: no CRIS, no fallback.
|
||||
id, with the data-residency implications of cross-region routing.
|
||||
|
||||
**GPT-5.5 / GPT-5.4 runtime availability** (model cards read 2026-11-06): US Geo CRIS (`us.openai.gpt-5.5` /
|
||||
`us.openai.gpt-5.4`) can be called from `us-east-1`, `us-east-2`, `us-west-1`, `us-west-2`, `ca-central-1` and
|
||||
`ca-west-1`. Global CRIS (`global.openai.gpt-5.5` / `global.openai.gpt-5.4`) is listed in those six plus
|
||||
`mx-central-1`, every EU region (`eu-central-1/2`, `eu-north-1`, `eu-south-1/2`, `eu-west-1/2/3`), `ap-east-2`,
|
||||
`ap-northeast-1/2/3`, `ap-south-1/2`, `ap-southeast-1` through `-7`, `il-central-1`, `me-central-1`, `me-south-1`,
|
||||
`af-south-1` and `sa-east-1` (33 regions). No runtime in-region form, no EU/India Geo profile, and no GovCloud
|
||||
runtime row. So the mantle matrix is no longer a hard gate for GPT-5.5 / GPT-5.4; only a strictly in-region
|
||||
requirement outside the mantle matrix blocks the same-model path.
|
||||
|
||||
Verify current footprints per model card / `aws___get_regional_availability` (AWS MCP Server) — the CRIS lists move faster than this file.
|
||||
|
||||
@@ -252,7 +273,8 @@ inference-option dimension:**
|
||||
workload has no data-residency constraint.
|
||||
|
||||
So the honest cost statement is conditional, not flat: a same-model GPT-5.6 move on Global CRIS is
|
||||
**cost-neutral**; the same move in-region or Geo (and any GPT-5.5 / GPT-5.4 move) is **~10% more expensive**. The
|
||||
**cost-neutral**; the same move in-region or Geo is **~10% more expensive**. The GPT-5.5 / GPT-5.4 cards publish
|
||||
In-Region pricing only (no Geo/Global rows as of 2026-11-06), so quote ~10% for them on every inference option. The
|
||||
same Global-CRIS-is-cost-neutral / in-Region-and-Geo-is-~10%-more pattern applies to the GPT-6 family below.
|
||||
Never state either number without stating the inference option it belongs to.
|
||||
|
||||
@@ -307,12 +329,15 @@ provider rate before claiming savings or parity for the selected context tier.
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier.
|
||||
|
||||
### GPT-5.5 / GPT-5.4 — In-Region only (no CRIS, no long-context tier; usable window 272K)
|
||||
### GPT-5.5 / GPT-5.4 — In-Region pricing only; short- and long-context tiers (model cards read 2026-11-06)
|
||||
|
||||
| Model | In-Region (in · out) | Cache read | Notes |
|
||||
| ------- | -------------------- | ---------- | -------------------------------- |
|
||||
| GPT-5.5 | 5.50 · 33.00 | 0.55 | no cache-write rate published |
|
||||
| GPT-5.4 | 2.75 · 16.50 | 0.275 | GovCloud (US-West): 3.30 · 19.80 |
|
||||
The cards publish only the In-Region inference option (the `us.` / `global.` runtime CRIS ids exist but carry no
|
||||
separate price row). Long-context rates apply to all input and output tokens of a request above 272K input tokens.
|
||||
|
||||
| Model | Short context ≤272K (in · cache read · out) | Long context >272K (in · cache read · out) | Notes |
|
||||
| ------- | ------------------------------------------- | ------------------------------------------ | ---------------------------------------- |
|
||||
| GPT-5.5 | 5.50 · 0.55 · 33.00 | 11.00 · 1.10 · 49.50 | no cache-write rate published |
|
||||
| GPT-5.4 | 2.75 · 0.275 · 16.50 | 5.50 · 0.55 · 24.75 | GovCloud (US-West) short: 3.375 · 0.3375 · 20.25 |
|
||||
|
||||
> **The Luna "blog discrepancy" resolved differently than first recorded.** The AWS News Blog's 0.20 / 1.20 is not
|
||||
> an error — it is the **Global CRIS** rate, now published on the Luna card. An earlier revision of this file said
|
||||
@@ -329,8 +354,10 @@ Its card does not establish the GPT-5.6 minimum-prefix, breakpoint, or quota-exe
|
||||
|
||||
Listed as a supported feature on the GPT-5.6 Sol/Terra/Luna model cards and on all five GPT-6 model cards (Astra,
|
||||
Sol, Luna, 6.1 Sol, 6.1 Terra — confirmed by the cache-write/cache-read columns in the GPT-6 pricing tables above). The
|
||||
GPT-5.5 and GPT-5.4 cards list client-side tool calling in that slot instead and do **not** list prompt caching.
|
||||
Do not assume caching on 5.5/5.4.
|
||||
GPT-5.5 and GPT-5.4 cards (read 2026-11-06) list **implicit prompt caching on `bedrock-runtime` Responses only**
|
||||
(automatic, no request parameters) with a cache-read rate and no cache-write rate; explicit caching is unsupported
|
||||
and the mantle feature list carries client-side tool calling in that slot instead. Do not assume caching on
|
||||
5.5/5.4 mantle, and do not inherit the GPT-5.6 prefix/breakpoint rules below for them.
|
||||
|
||||
| Property | Value |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
@@ -359,10 +386,10 @@ This corrects two claims that were previously applied to all Mantle traffic in t
|
||||
|
||||
- There is **no shared 10,000 RPM account limit** governing these models — the quota dimension is TPM, per model,
|
||||
per region.
|
||||
- "Switch to `bedrock-runtime`" is **not a throughput remedy for GPT-5.5 / GPT-5.4** — they have no
|
||||
`bedrock-runtime` path at all. For GPT-5.6 the runtime path DOES exist (CRIS only; see Access Paths above), so
|
||||
moving there is a legitimate option — but treat it as an endpoint/architecture choice with its own quota family
|
||||
and residency implications, not a free throughput escape hatch.
|
||||
- "Switch to `bedrock-runtime`" is **not a free throughput escape hatch**. For GPT-5.6, GPT-5.5 and GPT-5.4 the
|
||||
runtime path exists (CRIS only; see Access Paths above — GPT-5.5 / GPT-5.4 re-verified 2026-11-06), so moving
|
||||
there is a legitimate option — but treat it as an endpoint/architecture choice with its own quota family and
|
||||
residency implications.
|
||||
|
||||
The supported mitigations are: exponential backoff with a bounded retry count (`max_retries` on the OpenAI SDK),
|
||||
spreading load across minutes rather than bursting, ramping request rate gradually, and prompt caching (cached input
|
||||
@@ -409,8 +436,8 @@ This model family is moving fast (two GA waves and a repricing inside 10 weeks).
|
||||
|
||||
1. Re-read the [OpenAI model card index](https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html)
|
||||
for models added or removed, and each per-model card for lifecycle state and EOL date.
|
||||
2. Recheck the region matrix AND the GPT-5.6 CRIS footprints — for 5.5/5.4 a region change is migration-blocking;
|
||||
for 5.6 the runtime/CRIS path usually covers the region instead.
|
||||
2. Recheck the region matrix AND the CRIS footprints for every proprietary GPT model — the runtime/CRIS path
|
||||
usually covers a region outside the mantle matrix; only a strictly in-region requirement there is blocking.
|
||||
3. Recheck rates on the Bedrock pricing page OpenAI tab, and resolve any row still marked _unverified_.
|
||||
4. Recheck whether the Price List API has gained GPT-5.x coverage; if it has, drop the caveat above and let
|
||||
`estimate-ai.md` price these models from the MCP.
|
||||
@@ -424,3 +451,9 @@ This model family is moving fast (two GA waves and a repricing inside 10 weeks).
|
||||
2026), and GPT-6.1 Sol (GA Sep 29, 2026) are all confirmed on Bedrock with their own model cards and pricing
|
||||
tables. New discovery this refresh: a 6x-Standard **Ultrafast** service tier on Astra and 6.1 Sol only (Sol and
|
||||
Luna remain Standard-only) — the first new tier shape since this file started tracking the family.
|
||||
9. **2026-11-06 refresh — GPT-5.5 / GPT-5.4 re-read.** Item 5 fired again: both cards now list `bedrock-runtime`
|
||||
(Responses, Chat Completions, Converse via `us.` / `global.` CRIS ids; no in-region runtime form), Chat
|
||||
Completions on `bedrock-mantle`, a 1,050,000-token context window with 128,000 max output, a long-context
|
||||
(>272K) price tier, implicit prompt caching on runtime Responses, and a GovCloud (US-West) GPT-5.4 rate of
|
||||
3.375 · 20.25. Earlier "mantle-only / no CRIS / 272K / no long-context tier" claims for these two models were
|
||||
stale and are corrected throughout this file.
|
||||
|
||||
+4
-4
@@ -48,9 +48,9 @@ Inference on `bedrock-mantle` for these models is governed by **two per-model, p
|
||||
| Medium | Medium | Monitor 429s against **token** throughput, not request rate; enable prompt caching |
|
||||
| High | High | Enable prompt caching first (cached input is exempt from input TPM), then request a quota increase |
|
||||
|
||||
**The `bedrock-runtime` fallback exists only for GPT-5.6 and GPT-6 (Astra, Sol, Luna, 6.1 Sol, 6.1 Terra).** GPT-5.5 and GPT-5.4 are `bedrock-mantle` only and in-region only — for them, "switch to `bedrock-runtime`" requires moving to a different model (Bedrock-native or `gpt-oss`), a model change with its own eval cost. GPT-5.6 Sol/Terra/Luna and GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra DO have a `bedrock-runtime` path via CRIS inference profiles (GPT-5.6 `us.`/`in.`/`global.`, GPT-6 `us.`/`global.` prefixed ids; the model cards recommend runtime for new applications) — a legitimate endpoint option with its own quota family, and on Global CRIS it is also the cost-parity option. Scaling levers on the mantle path, in order:
|
||||
**Every proprietary GPT model on Bedrock has a `bedrock-runtime` path via CRIS inference profiles.** GPT-5.6 Sol/Terra/Luna (`us.`/`in.`/`global.`), GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra (`us.`/`global.`), and GPT-5.5 / GPT-5.4 (`us.`/`global.`; per their model cards read 2026-11-06, [GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) all list `bedrock-runtime` with Responses, Chat Completions and Converse; none has a runtime in-region form. The model cards recommend runtime for new applications. It is a legitimate endpoint option with its own quota family, and for GPT-5.6 / GPT-6 on Global CRIS it is also the cost-parity option (the GPT-5.5 / GPT-5.4 cards publish In-Region pricing only). A same-model endpoint switch is not a model change; only moving to a Bedrock-native model or `gpt-oss` carries the eval cost of a model change. Scaling levers on the mantle path, in order:
|
||||
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 only) — 90% off cached input and exempt from the input-TPM quota
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 on mantle; GPT-5.5 / GPT-5.4 list implicit caching on `bedrock-runtime` Responses only) — 90% off cached input and exempt from the input-TPM quota
|
||||
2. Exponential backoff with a bounded retry count (`max_retries` on the OpenAI SDK)
|
||||
3. Spreading load across minutes and ramping request rate gradually rather than bursting
|
||||
4. A Service Quotas increase for that model's input/output TPM in that region
|
||||
@@ -108,8 +108,8 @@ Claude models on Mantle have an additional **output TPM cap** that differs by mo
|
||||
|
||||
- For Claude migrations at medium/high volume: the 2M output TPM cap on Claude 4.7+ is the binding constraint
|
||||
- For OpenAI proprietary GPT targets: this Claude cap does not apply. Use the GPT-5.x mantle quotas or
|
||||
Astra-specific verification above. Astra / GPT-5.6 have runtime CRIS alternatives with their own
|
||||
quotas; GPT-5.5/5.4 have no runtime path
|
||||
Astra-specific verification above. Astra / GPT-5.6 / GPT-5.5 / GPT-5.4 all have runtime CRIS alternatives
|
||||
with their own quotas (`us.`/`global.` prefixed ids; no runtime in-region form)
|
||||
- For `gpt-oss` targets: these do run on `bedrock-runtime`, so standard account TPM limits and the Converse-path mitigations apply
|
||||
- When output-heavy workloads (long JSON, tool outputs, multi-step reasoning) are detected, flag the relevant cap prominently; recommend `bedrock-runtime` for production **only** when the target model actually has a `bedrock-runtime` path
|
||||
|
||||
|
||||
+32
-25
@@ -42,7 +42,9 @@ Two consequences that invert the previous decision logic:
|
||||
2. **The cost of a same-model move depends on the inference option — state it conditionally.** GPT-5.x In-Region and Geo
|
||||
CRIS are priced at OpenAI's _data residency_ tier, exactly 1.10x the standard list price, so those moves cost
|
||||
~10% more. **GPT-5.6 on Global CRIS is priced at the standard list price — cost parity** — available only when
|
||||
the workload has no data-residency constraint (GPT-5.5 / GPT-5.4 have no CRIS at all). See
|
||||
the workload has no data-residency constraint. GPT-5.5 / GPT-5.4 also have `us.`/`global.` CRIS ids on
|
||||
`bedrock-runtime`, but their model cards (read 2026-11-06) publish In-Region pricing only, so quote the ~10%
|
||||
figure for them on every inference option until a Global rate is published. See
|
||||
`shared/openai-on-bedrock.md`. The non-cost case — AWS commitments, IAM/VPC/CloudTrail governance, residency,
|
||||
prompt caching, one vendor relationship — carries the in-region path; never present that path as free or neutral.
|
||||
For Astra, use its separately dated Bedrock rates and verify the source provider's rate before comparing.
|
||||
@@ -72,14 +74,15 @@ GPT-5.4, the target is **the same model on Bedrock**.
|
||||
| GPT-5.6 Sol | GPT-5.6 Sol | `openai.gpt-5.6-sol` / `us.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Terra | GPT-5.6 Terra | `openai.gpt-5.6-terra` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Luna | GPT-5.6 Luna | `openai.gpt-5.6-luna` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
|
||||
Then apply the **region gate** below. Record `model_change: false` and the chosen path in
|
||||
`aws-design-ai.json` → `ai_architecture.code_migration`: `migration_path: "mantle_openai_chat"` for
|
||||
Astra sources using Chat Completions, or `"mantle_openai_responses"` for a selected Mantle Responses
|
||||
path. Use `migration_path: "runtime_openai_cris"` when a GPT-6 / GPT-5.6 target is served via
|
||||
`bedrock-runtime` with a CRIS id (the model cards recommend runtime for new applications).
|
||||
sources using Chat Completions on a model whose card lists it on mantle (GPT-6, GPT-5.5, GPT-5.4), or
|
||||
`"mantle_openai_responses"` for a selected Mantle Responses path. Use `migration_path: "runtime_openai_cris"`
|
||||
when a GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 target is served via `bedrock-runtime` with a CRIS id (the model
|
||||
cards recommend runtime for new applications).
|
||||
|
||||
Report the assessment honestly: `strong_migrate` here means "low-risk, well-supported move", not "cheaper".
|
||||
|
||||
@@ -105,14 +108,16 @@ regions — offer it, stating the data-residency implication of cross-region rou
|
||||
Global routes anywhere and is the cost-parity option). Only a workload that requires strictly in-region processing
|
||||
in a region outside the mantle matrix falls through to Tier 1.
|
||||
|
||||
**GPT-5.5 / GPT-5.4:** hard gate. Mantle in-region only, no CRIS, no fallback. If the target region is not in the
|
||||
matrix:
|
||||
**GPT-5.5 / GPT-5.4:** same shape as GPT-5.6, with a narrower mantle matrix (GPT-5.5: us-east-1, us-east-2;
|
||||
GPT-5.4: us-east-1, us-east-2, us-west-2, us-gov-west-1 — per their model cards, read 2026-11-06). If the target
|
||||
region is in that matrix, either endpoint works. Otherwise the `bedrock-runtime` CRIS path (`us.openai.gpt-5.5` /
|
||||
`global.openai.gpt-5.5`, `us.openai.gpt-5.4` / `global.openai.gpt-5.4`) reaches it, with the same data-residency
|
||||
caveat as GPT-5.6 (Geo stays within the geography, Global routes anywhere). There is no runtime in-region form.
|
||||
Only a workload that requires strictly in-region processing in a region outside the mantle matrix falls through:
|
||||
|
||||
1. Prefer offering a **region change** to a supported region, if no data-residency constraint pins the workload.
|
||||
This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. If the region is fixed, also offer **upgrading within the vendor to GPT-5.6** (which reaches the region via
|
||||
CRIS) alongside the cross-family option — a generation change, but same vendor and prompt idioms.
|
||||
3. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
1. Prefer offering a **region change** to a supported mantle region, if no data-residency constraint pins the
|
||||
workload. This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
|
||||
Record the outcome in `regional_warnings[]`.
|
||||
|
||||
@@ -235,10 +240,11 @@ model on price.
|
||||
|
||||
### Open-weight gpt-oss
|
||||
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and, unlike the
|
||||
proprietary GPT models, **do** support `bedrock-runtime` / Converse. They sit a capability class below the GPT-5.x
|
||||
frontier tier. Recommend them when the user wants OpenAI-architecture models on the Bedrock-native runtime surface
|
||||
(for Guardrails, invocation logging, or Converse-based tooling) rather than through mantle.
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and support
|
||||
`bedrock-runtime` / Converse with a bare model ID (no CRIS profile needed) and `Invoke`, which the proprietary GPT
|
||||
models do not. They sit a capability class below the GPT-5.x frontier tier. Recommend them when the user wants
|
||||
OpenAI-architecture models on the Bedrock-native runtime surface (for Guardrails, invocation logging, or
|
||||
Converse-based tooling) at the lowest price rather than through mantle.
|
||||
|
||||
---
|
||||
|
||||
@@ -284,7 +290,7 @@ rather than dismissing it. For a long-context (>272K) workload the gap is much w
|
||||
| Chat Completions | Preserve supported Chat when selected | Reshape only for a selected API change; verify the model and endpoint |
|
||||
| Function calling | Supported on GPT-5.x via mantle; Claude tools via Converse | Same-model keeps tool semantics identical |
|
||||
| Reasoning effort | `reasoning={"effort": ...}` on mantle | `none` / `low` / `medium` / `high` / `xhigh` / `max` |
|
||||
| Prompt caching | GPT-5.6 only (90% off cached input) | Not listed for GPT-5.5 / 5.4; Claude has its own caching |
|
||||
| Prompt caching | GPT-5.6 (90% off cached input); GPT-5.5 / 5.4 implicit caching on `bedrock-runtime` Responses only | GPT-5.5 / 5.4 cards (read 2026-11-06) list a cache-read rate and implicit caching on runtime, none on mantle; Claude has its own caching |
|
||||
| Streaming | Supported | Verify per surface |
|
||||
| Vision | GPT-5.x (image input) or Claude / Llama 4 | Same-model preserves behavior |
|
||||
| JSON mode | Claude (excellent), Nova Pro (good) | Most models via prompt |
|
||||
@@ -339,11 +345,12 @@ and closed to new customers as of July 30, 2026.
|
||||
### When to prefer a Bedrock-native model over the mantle GPT path
|
||||
|
||||
Choose Claude / Nova via Converse when the workload needs Knowledge Bases, intelligent prompt routing, structured
|
||||
outputs, or application inference profiles — none of which the GPT path offers. Note what no longer forces a family
|
||||
switch for GPT-5.6: its `bedrock-runtime` CRIS path provides cross-region inference, invocation logs, and
|
||||
Guardrails (Converse API only; prompt caching is Responses-only on runtime). For GPT-5.5 / GPT-5.4, Guardrails,
|
||||
logging, and cross-region reach still require a different model. The mantle GPT path is the smallest change; the
|
||||
Bedrock-native path remains the most feature-complete.
|
||||
outputs, or application inference profiles — none of which the mantle GPT path offers. Note what no longer forces
|
||||
a family switch for GPT-5.6, GPT-5.5 and GPT-5.4: their `bedrock-runtime` CRIS paths provide cross-region
|
||||
inference, invocation logs, Guardrails and application inference profiles (both Converse API only), and
|
||||
structured outputs and implicit prompt caching (both Responses API only) — per the GPT-5.4 / GPT-5.5 model cards
|
||||
read 2026-11-06. Knowledge Bases and intelligent prompt routing still require a Bedrock-native model. The mantle
|
||||
GPT path is the smallest change; the Bedrock-native path remains the most feature-complete.
|
||||
|
||||
### High-spend tiering
|
||||
|
||||
@@ -364,8 +371,8 @@ alternative, not against the source provider.
|
||||
cheaper and exempt from the input-TPM quota, which relieves the binding constraint on mantle.
|
||||
|
||||
**Very high (>100M tokens/day):** tiering plus caching. Check TPM headroom per model per region early; there is no
|
||||
RPM quota and no `bedrock-runtime` fallback for GPT models, so quota increases are the scaling lever. See
|
||||
`shared/ai-migration-guardrails.md`.
|
||||
RPM quota, and the `bedrock-runtime` CRIS path for GPT-5.6 / 5.5 / 5.4 / GPT-6 carries its own quotas rather than
|
||||
adding headroom to mantle, so quota increases are the scaling lever. See `shared/ai-migration-guardrails.md`.
|
||||
|
||||
---
|
||||
|
||||
|
||||
+20
-12
@@ -133,8 +133,8 @@ See `vendored/ai/ai-model-lifecycle.md` for lifecycle details. **Do not recommen
|
||||
| GPT-5.6 Sol | openai.gpt-5.6-sol | OpenAI | 4.40 | 22.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 8.80/33.00) |
|
||||
| GPT-5.6 Terra | openai.gpt-5.6-terra | OpenAI | 2.20 | 13.20 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 4.40/19.80) |
|
||||
| GPT-5.6 Luna | openai.gpt-5.6-luna | OpenAI | 0.22 | 1.32 | 272K | fast | active (mantle in-region + runtime CRIS; 1M tier 0.44/1.98) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle only; no 1M tier) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle only; no 1M tier) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 11.00/49.50; card read 2026-11-06) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 5.50/24.75; card read 2026-11-06) |
|
||||
| Gemma 4 31B | google.gemma-4-31b | Google | 0.14 | 0.40 | 256K | mid | active (Mantle only) |
|
||||
| Gemma 4 26B-A4B | google.gemma-4-26b-a4b | Google | 0.13 | 0.40 | 256K | efficient | active (Mantle only) |
|
||||
| Gemma 4 E2B | google.gemma-4-e2b | Google | 0.04 | 0.08 | 128K | budget | active (Mantle only) |
|
||||
@@ -353,17 +353,24 @@ Regions (mantle): us-east-1 / us-east-2 for all five; us-west-2 additionally car
|
||||
| GPT-5.5 | 5.50 | 33.00 | — | 0.55 |
|
||||
| GPT-5.4 | 2.75 | 16.50 | — | 0.275 |
|
||||
|
||||
#### Long context (1M) — GPT-5.6 only, 2.0x input / 1.5x output
|
||||
#### Long context (>272K input tokens) — GPT-5.6 at 2.0x input / 1.5x output; GPT-5.5 / GPT-5.4 per their model cards
|
||||
|
||||
| Model | Input $/1M | Output $/1M |
|
||||
| ------------- | ---------- | ----------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 |
|
||||
| Model | Input $/1M | Output $/1M | Cache read $/1M |
|
||||
| ------------- | ---------- | ----------- | --------------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 | — |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 | — |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 | — |
|
||||
| GPT-5.5 | 11.00 | 49.50 | 1.10 |
|
||||
| GPT-5.4 | 5.50 | 24.75 | 0.55 |
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier. GPT-5.5 and GPT-5.4 have no such tier, so
|
||||
their usable window is 272K. Global CRIS is published at OpenAI's standard list price (cost parity, GPT-5.6 only —
|
||||
see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
A workload above 272K input tokens must be priced at the long-context tier, applied to all input and output tokens
|
||||
of the request. GPT-5.5 and GPT-5.4 rows read from their model cards on 2026-11-06
|
||||
([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html),
|
||||
[GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)); both cards state a
|
||||
1,050,000-token context window (128,000 max output) and list `us.`/`global.` CRIS ids on `bedrock-runtime`, but
|
||||
publish the In-Region inference option only. Global CRIS is published at OpenAI's standard list price (cost parity,
|
||||
GPT-5.6 only — see `openai-on-bedrock.md`). GovCloud (US-West) GPT-5.4 short-context is 3.375 / 20.25 (cache read
|
||||
0.3375) per the 2026-11-06 card; an earlier revision recorded 3.30 / 19.80.
|
||||
|
||||
> **One source conflict.** These models are absent
|
||||
> from the AWS Price List API entirely: querying `AmazonBedrock` returns only `gpt-oss` and GPT OSS Safeguard, and
|
||||
@@ -373,7 +380,8 @@ see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
|
||||
**Prompt caching (GPT-5.6 only):** cached input read at a 90% discount, cache write at 1.25x uncached input, minimum
|
||||
1,024-token prefix, up to 4 breakpoints, ≥30-minute retention. Cached input is exempt from the input-TPM quota.
|
||||
GPT-5.5 and GPT-5.4 do not list prompt caching support.
|
||||
GPT-5.5 and GPT-5.4 list only implicit prompt caching on `bedrock-runtime` Responses (automatic, no request
|
||||
parameters; cache-read rate above, no cache-write rate) and none on mantle (model cards read 2026-11-06).
|
||||
|
||||
#### Open-weight gpt-oss
|
||||
|
||||
|
||||
+15
-8
@@ -190,12 +190,18 @@ Traditional-AI capabilities (AWS AI service, not Bedrock): `document_extraction`
|
||||
## migration_path vocabulary
|
||||
|
||||
`mantle_openai_responses` (an OpenAI-protocol source — direct or Azure-fronted — with the model
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `converse`
|
||||
(Bedrock-native Converse/`bedrock-runtime`; set `model_change: true` when a proprietary GPT
|
||||
model must move off mantle for Guardrails / Knowledge Bases / logging / an unsupported region) ·
|
||||
`gpt-oss` (OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic
|
||||
Bedrock SDK swap). A proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — those models are mantle-only.
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `mantle_openai_chat`
|
||||
(same, preserving a Chat Completions source on a model whose card lists Chat Completions on
|
||||
mantle — GPT-6, GPT-5.5, GPT-5.4) · `runtime_openai_cris` (same-model move served via
|
||||
`bedrock-runtime` with a `us.`/`in.`/`global.` CRIS id, `model_change: false`; available for
|
||||
GPT-6, GPT-5.6, GPT-5.5 and GPT-5.4 per their model cards) · `converse` (Bedrock-native
|
||||
Converse/`bedrock-runtime`; set `model_change: true` when the workload needs Knowledge Bases,
|
||||
intelligent prompt routing, or another feature no same-model path serves) · `gpt-oss`
|
||||
(OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic Bedrock SDK
|
||||
swap). A **bare** proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — bare ids are mantle ids; every proprietary GPT target on
|
||||
`bedrock-runtime` carries a CRIS-prefixed id (there is no runtime in-region form for GPT-6,
|
||||
GPT-5.6, GPT-5.5 or GPT-5.4).
|
||||
|
||||
## regional_validation
|
||||
|
||||
@@ -255,8 +261,9 @@ rubric file (`ai.md`, a provider-to-Bedrock mapping guide, or a `vendored/ai/*`
|
||||
`price_comparison`.
|
||||
- [ ] `honest_assessment` overall is the weakest across `bedrock_models[]`; if
|
||||
`recommend_stay`, `honest_assessment_reason` names a non-cost blocker.
|
||||
- [ ] No `bedrock_models[]` entry pairs a proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (mantle-only).
|
||||
- [ ] No `bedrock_models[]` entry pairs a **bare** proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (bare ids are mantle ids; a `bedrock-runtime`
|
||||
target for GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 carries a `us.`/`in.`/`global.` CRIS id).
|
||||
- [ ] For a retained source whose `source_service` is `azure_openai`, or whose provider resolves
|
||||
to direct OpenAI: every source model that is available on Bedrock and carried by the
|
||||
target region maps to ITSELF with `model_change: false` — not to a Claude/Nova substitute.
|
||||
|
||||
@@ -1189,7 +1189,7 @@ Tell the user, before collecting models:
|
||||
> - **OpenAI on Bedrock comes in a few forms, all real** — and which endpoint they use is per-ID, not one blanket rule:
|
||||
> - *Open-weight* `gpt-oss` (`openai.gpt-oss-20b-1:0`, `openai.gpt-oss-120b-1:0`) → `bedrock-runtime` via `bedrock:InvokeModel` / `Converse` (its Responses API is also offered on `bedrock-mantle`).
|
||||
> - *Bare proprietary* GPT ids (`openai.gpt-5*`, not `gpt-oss`) → probed on the `bedrock-mantle` endpoint (Responses API).
|
||||
> - A *GPT-5.6 CRIS profile* id prefixed `us.` / `in.` / `global.` (e.g. `global.openai.gpt-5.6-sol`) → a `bedrock-runtime` target where Converse is supported.
|
||||
> - A *CRIS profile* id prefixed `us.` / `in.` / `global.` (e.g. `global.openai.gpt-5.6-sol`, `us.openai.gpt-5.4`, `global.openai.gpt-6-astra`) → a `bedrock-runtime` target where Converse is supported; every proprietary GPT model on Bedrock (GPT-5.6, GPT-6, GPT-5.5, GPT-5.4) has this form, and none has a runtime in-region form.
|
||||
> - The `bedrock-mantle` path uses a **separate** action set — `bedrock-mantle:*` (e.g. the `AmazonBedrockMantleInferenceAccess` managed policy: `bedrock-mantle:CreateInference` + `CallWithBearerToken`) — distinct from `bedrock:InvokeModel`. The preflight decides per model; don't assume.
|
||||
> - What is **not** on Bedrock is calling OpenAI's own hosted API at api.openai.com — that stays with OpenAI. 'GPT on Bedrock' means the AWS-served models above, reached through AWS endpoints and IAM.
|
||||
>
|
||||
@@ -1277,8 +1277,9 @@ Then, per B4's branch table:
|
||||
after they enable it. This is the common, expected outcome for a user who came here to "get
|
||||
access."
|
||||
- `reason: authz` → access is enabled but IAM denies inference. Name the action to grant — for a
|
||||
standard model `bedrock:InvokeModel`; for a mantle-only `openai.gpt-5*` target (bare proprietary
|
||||
GPT, not `gpt-oss`) the `bedrock-mantle:*` set (see **B4a**). The `detail` says which — follow it
|
||||
standard model `bedrock:InvokeModel`; for a bare `openai.gpt-5*` / `openai.gpt-6*` mantle target
|
||||
(bare proprietary GPT, not `gpt-oss`, not a `us.`/`in.`/`global.` CRIS id) the `bedrock-mantle:*`
|
||||
set (see **B4a**). The `detail` says which — follow it
|
||||
rather than guessing from the ID, since gpt-oss fails on `bedrock:InvokeModel`, not mantle.
|
||||
- `reason: model_unavailable` → the ID isn't offered in `$REGION`. Use the
|
||||
`resolve-bedrock-model-id` procedure to suggest a cross-region inference-profile ID or a
|
||||
|
||||
+2
-2
@@ -29,7 +29,7 @@ for OpenAI-compatible runtime calls, CRIS ids, and runtime IAM permissions. Astr
|
||||
- `resolution_kind`: `ux_choice`
|
||||
- Source (OpenAI): `client.chat.completions.create(...)`, `messages=[...]`, reads `choices[0].message.content`
|
||||
- Target (mantle GPT): `client.responses.create(...)`, `input=...`, reads `output_text`
|
||||
- Chat Completions support is **model-specific, not uniformly unverified.** All four GPT-6 models (Astra, Sol, Luna, 6.1 Sol) and the GPT-5.6 family (Sol, Terra, Luna) list `Chat Completions` as a supported API on both `bedrock-mantle` and `bedrock-runtime` CRIS (per their model cards — see `openai-on-bedrock.md` § Access Paths). For these models, a Chat Completions source MAY keep `chat.completions.create` and `choices[0].message` parsing unchanged — **reshape to Responses is OPTIONAL**, driven only by a Responses-only feature (runtime prompt caching, reasoning-item round-trips on `bedrock-runtime`) or a demonstrated account-specific constraint, not by a blanket assumption that Chat Completions is unsupported. For GPT-5.5 / GPT-5.4 (mantle-only, no CRIS) Chat Completions support is still unconfirmed on their model cards — treat those two as requiring a reshape to Responses, and probe the target account before committing.
|
||||
- Chat Completions support is **model-specific, not uniformly unverified.** All four GPT-6 models (Astra, Sol, Luna, 6.1 Sol) and the GPT-5.6 family (Sol, Terra, Luna) list `Chat Completions` as a supported API on both `bedrock-mantle` and `bedrock-runtime` CRIS (per their model cards — see `openai-on-bedrock.md` § Access Paths). For these models, a Chat Completions source MAY keep `chat.completions.create` and `choices[0].message` parsing unchanged — **reshape to Responses is OPTIONAL**, driven only by a Responses-only feature (runtime prompt caching, reasoning-item round-trips on `bedrock-runtime`) or a demonstrated account-specific constraint, not by a blanket assumption that Chat Completions is unsupported. GPT-5.5 / GPT-5.4 follow the same rule: their model cards (read 2026-11-06 — [GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) list `Chat Completions` as supported on both `bedrock-mantle` and `bedrock-runtime` (`us.`/`global.` CRIS ids), so the reshape is OPTIONAL for them too. Plan a reshape only when the selected model's card leaves Chat Completions unchecked on the selected endpoint, and probe the target account before committing, as for any model.
|
||||
|
||||
### detect_grep
|
||||
|
||||
@@ -66,7 +66,7 @@ grep -rEn 'function_call_output|tool_call' <REPO> --include="*.py" --include="*.
|
||||
## prompt-caching-availability
|
||||
|
||||
- `resolution_kind`: `impl_path`
|
||||
- Prompt caching is listed as supported on GPT-5.6 and GPT-6 only. Do not emit caching configuration for `openai.gpt-5.5` or `openai.gpt-5.4`.
|
||||
- Prompt caching with request parameters (explicit mode) is listed on GPT-5.6 and GPT-6 only. The GPT-5.5 / GPT-5.4 cards (read 2026-11-06) list only implicit caching on `bedrock-runtime` Responses (automatic, no request parameters) and none on mantle — do not emit caching configuration for `openai.gpt-5.5` or `openai.gpt-5.4` on either endpoint.
|
||||
|
||||
- Astra lists implicit and explicit prompt caching on mantle Responses, with **30-minute cache-write** rates.
|
||||
Do not inherit GPT-5.6's minimum-prefix, breakpoint, or cached-input quota-exemption rules.
|
||||
|
||||
+23
-16
@@ -50,9 +50,13 @@ aws bedrock list-foundation-models \
|
||||
**Check this before Step 1.** Route by family and endpoint (GPT-5 verified 2026-08-21; Astra regions verified 2026-09-30; see
|
||||
`gcp-to-aws/references/shared/openai-on-bedrock.md`):
|
||||
|
||||
**Case A — GPT-5.5 / GPT-5.4 (`openai.gpt-5.5`, `openai.gpt-5.4`): mantle-only, no inference profile.** The
|
||||
inference-profile path below cannot resolve them — `list-inference-profiles` never returns them, and Step 3's token
|
||||
ranking would end in a spurious `blocked` for a perfectly valid id. Validate against the model catalog instead:
|
||||
**Case A — bare GPT-5.5 / GPT-5.4 ids (`openai.gpt-5.5`, `openai.gpt-5.4`) when the plan targets the mantle
|
||||
endpoint: bare mantle id, not an inference profile.** The inference-profile path below cannot resolve the bare form —
|
||||
`list-inference-profiles` returns only the CRIS-prefixed ids (`us.openai.gpt-5.5`, `global.openai.gpt-5.5`,
|
||||
`us.openai.gpt-5.4`, `global.openai.gpt-5.4`, per the model cards read 2026-11-06), never the bare mantle id — and
|
||||
Step 3's token ranking would end in a spurious `blocked` for a perfectly valid id. (When the plan targets
|
||||
`bedrock-runtime` for GPT-5.5 / GPT-5.4, the CRIS id flows through Case B and Steps 1–5 like GPT-5.6.) Validate the
|
||||
bare id against the model catalog instead:
|
||||
|
||||
```bash
|
||||
aws bedrock list-foundation-models \
|
||||
@@ -69,7 +73,7 @@ aws bedrock list-foundation-models \
|
||||
this catalog's `modelId` + `modelName` pairs instead of against inference profiles. **Do this check before
|
||||
falling through to Step 1** — a friendly name like `"OpenAI GPT-5.5"` (tokens `{openai, gpt, 5}`) has full token
|
||||
overlap with `openai.gpt-5.5`'s own id (`{openai, gpt, 5}`) and would otherwise fall through to Step 1's
|
||||
`list-inference-profiles`, which never lists a mantle-only model — the free-text input would then reach Step 3's
|
||||
`list-inference-profiles`, which never lists a bare mantle id — the free-text input would then reach Step 3's
|
||||
ranking with zero candidates and dead-end at `blocked` with no matches shown, even though the foundation catalog
|
||||
had the exact model the whole time. A single candidate with full token overlap on the model-identifying tokens
|
||||
(excluding generic tokens like a bare `gpt`) → surface it as the ONE candidate through Step 4 exactly as an
|
||||
@@ -78,24 +82,27 @@ aws bedrock list-foundation-models \
|
||||
list alongside any inference-profile candidates Steps 1–3 separately produce; do not silently drop either source.
|
||||
- **No match at all (exact or token) and `plan_model_id` looked ID-shaped, not free-text** → not enabled or not
|
||||
available in this region. Return `blocked` with `reason: model_unresolvable`, putting the region and the
|
||||
`openai.*` ids that _were_ returned in `detail`. These two models have no CRIS, so the remedy is a region change
|
||||
or a different model — never an inference-profile prefix.
|
||||
`openai.*` ids that _were_ returned in `detail`. Do not silently add an inference-profile prefix to rescue a
|
||||
mantle plan: the remedy is a region change, a different model, or a plan change to the `bedrock-runtime` CRIS
|
||||
path (`us.` / `global.` ids, Case B) — a decision for the plan, not for this resolver.
|
||||
- CLI failure / missing `bedrock:ListFoundationModels` → `blocked` with `reason: model_unresolvable` and the exact
|
||||
error in `detail`, rather than guessing.
|
||||
|
||||
These two need `bedrock-mantle:*` IAM actions (e.g. `AmazonBedrockMantleInferenceAccess`), not
|
||||
The bare ids need `bedrock-mantle:*` IAM actions (e.g. `AmazonBedrockMantleInferenceAccess`), not
|
||||
`bedrock:InvokeModel`; resolution success does not imply invoke authorization.
|
||||
|
||||
**A free-text name that does NOT token-match this catalog** (e.g. `"anthropic claude haiku"`, `"Nova Micro"`)
|
||||
continues to Step 1 as before — this catalog check only short-circuits the specific case a mantle-only model would
|
||||
continues to Step 1 as before — this catalog check only short-circuits the specific case a bare mantle id would
|
||||
otherwise be unreachable from free text; it is not a replacement for Step 1's inference-profile search for every
|
||||
other provider/model.
|
||||
|
||||
**Case B — GPT-5.6 and GPT-6 (`openai.gpt-5.6-sol` / `-terra` / `-luna`, `openai.gpt-6-astra` / `-sol` / `-luna`,
|
||||
`openai.gpt-6.1-sol`, or already `us.` / `in.` / `global.` prefixed): BOTH paths exist.** The bare id is the mantle
|
||||
form; `bedrock-runtime` serves these models through CRIS inference profiles (`us.openai.gpt-5.6-*` /
|
||||
`us.openai.gpt-6-*`, `in.openai.gpt-5.6-*` in India Regions, `global.openai.gpt-5.6-*` / `global.openai.gpt-6-*`),
|
||||
which `list-inference-profiles` DOES return. Route on the plan's intent:
|
||||
**Case B — GPT-5.6, GPT-6, and GPT-5.5 / GPT-5.4 on `bedrock-runtime` (`openai.gpt-5.6-sol` / `-terra` / `-luna`,
|
||||
`openai.gpt-6-astra` / `-sol` / `-luna`, `openai.gpt-6.1-sol`, `openai.gpt-5.5` / `openai.gpt-5.4` with a runtime
|
||||
plan, or already `us.` / `in.` / `global.` prefixed): BOTH paths exist.** The bare id is the mantle form;
|
||||
`bedrock-runtime` serves these models through CRIS inference profiles (`us.openai.gpt-5.6-*` / `us.openai.gpt-6-*` /
|
||||
`us.openai.gpt-5.5` / `us.openai.gpt-5.4`, `in.openai.gpt-5.6-*` in India Regions, `global.openai.gpt-5.6-*` /
|
||||
`global.openai.gpt-6-*` / `global.openai.gpt-5.5` / `global.openai.gpt-5.4`), which `list-inference-profiles` DOES
|
||||
return. Route on the plan's intent:
|
||||
|
||||
- Plan targets the mantle endpoint (`migration_path` starts with `mantle`, or the id is bare) → validate the bare id
|
||||
against the model catalog exactly as in Case A.
|
||||
@@ -217,9 +224,9 @@ stops on abort.
|
||||
- This skill is idempotent: calling it twice with the same already-validated
|
||||
ID will hit Step 0 (documented bare-ID route) or Step 2 (inference profile) and return immediately.
|
||||
- Steps 1–5 assume the target is a `bedrock-runtime` model reachable through an
|
||||
inference profile. Mantle-only ids (GPT-5.5/5.4, and GPT-5.6 when the plan
|
||||
targets the mantle endpoint), and London in-region Opus 5.5, are handled entirely in Step 0; GPT-5.6/GPT-6 CRIS ids
|
||||
flow through Steps 1–5 like any other inference profile. See
|
||||
inference profile. Bare mantle ids (any proprietary GPT model — GPT-5.5/5.4, GPT-5.6, GPT-6 — when the plan
|
||||
targets the mantle endpoint), and London in-region Opus 5.5, are handled entirely in Step 0; GPT-5.6/GPT-6 and
|
||||
GPT-5.5/5.4 CRIS ids flow through Steps 1–5 like any other inference profile. See
|
||||
`$ASSESS_BASE/references/shared/openai-on-bedrock.md` for the authoritative
|
||||
family split.
|
||||
- Output of this skill should replace the plan's `target_model_id` in the
|
||||
|
||||
@@ -90,7 +90,11 @@ STATIC_FALLBACK = {
|
||||
# price — cost PARITY, for workloads with no residency constraint
|
||||
# The GPT-5.6 family also has a LONG-CONTEXT (1M) tier at 2.0x input / 1.5x
|
||||
# output per option, NOT represented here — a >272K workload priced from this
|
||||
# table is understated. GPT-5.5 and GPT-5.4: mantle-only, no CRIS, no 1M tier.
|
||||
# table is understated. GPT-5.5 and GPT-5.4 (model cards read 2026-11-06) also
|
||||
# have a long-context (>272K) tier (5.5: 11.00/49.50; 5.4: 5.50/24.75 per 1M),
|
||||
# likewise NOT represented here. Their cards publish the In-Region option only;
|
||||
# the us./global. CRIS ids they list on bedrock-runtime carry no separate price
|
||||
# row, so no prefixed GPT-5.5/5.4 entries are added below until one is published.
|
||||
"openai.gpt-5.6-sol": {"input_per_1k_usd": 0.0044, "output_per_1k_usd": 0.022},
|
||||
"openai.gpt-5.6-terra": {"input_per_1k_usd": 0.0022, "output_per_1k_usd": 0.0132},
|
||||
"openai.gpt-5.6-luna": {"input_per_1k_usd": 0.00022, "output_per_1k_usd": 0.00132},
|
||||
@@ -135,9 +139,12 @@ STATIC_FALLBACK = {
|
||||
|
||||
|
||||
def is_mantle_gpt(model_id: str) -> bool:
|
||||
"""Pure: OpenAI's proprietary GPT models (GPT-5.5/5.4/5.6 and GPT-6
|
||||
Astra/Sol/Luna/6.1 Sol), which the AWS PriceList API does not carry. The
|
||||
open-weight gpt-oss models ARE in the PriceList API and must not match."""
|
||||
"""Pure: BARE ids of OpenAI's proprietary GPT models (GPT-5.5/5.4/5.6 and GPT-6
|
||||
Astra/Sol/Luna/6.1 Sol) — the bedrock-mantle id form — which the AWS PriceList
|
||||
API does not carry. The `us.`/`in.`/`global.` CRIS forms (bedrock-runtime; GPT-5.5
|
||||
and 5.4 have `us.`/`global.` ids too) do not match here; callers pair this with
|
||||
an explicit prefixed-id regex. The open-weight gpt-oss models ARE in the
|
||||
PriceList API and must not match."""
|
||||
mid = model_id.lower()
|
||||
return (mid.startswith("openai.gpt-5") or mid.startswith("openai.gpt-6")) and "oss" not in mid
|
||||
|
||||
|
||||
@@ -20,12 +20,14 @@ def is_inference_profile(model_id: str) -> bool:
|
||||
|
||||
def is_mantle_model(model_id: str, region: str = "") -> bool:
|
||||
"""True for bare proprietary GPT-5/GPT-6 IDs and the bare Opus 5.5 ID, which use
|
||||
bedrock-mantle (Responses and Messages respectively). They need
|
||||
bedrock-mantle (Responses/Chat Completions and Messages respectively). They need
|
||||
`bedrock-mantle:*` actions — a policy
|
||||
granting only `bedrock:InvokeModel` against a foundation-model ARN cannot
|
||||
authorize these bare IDs. Opus 5.5 in eu-west-2 is a runtime in-region
|
||||
target (model card verified 2026-10-03); its other bare-ID regions use Mantle.
|
||||
Their prefixed forms use runtime instead.
|
||||
The `us.`/`in.`/`global.` prefixed forms use bedrock-runtime instead — for every
|
||||
proprietary GPT model, GPT-5.5/5.4 included (their model cards list `us.`/`global.`
|
||||
CRIS ids; none has a runtime in-region form), so a prefixed id must NOT match.
|
||||
The open-weight gpt-oss models DO use bedrock-runtime and must not match."""
|
||||
mid = model_id.lower()
|
||||
return ((mid.startswith("openai.gpt-5") or mid.startswith("openai.gpt-6")) and "oss" not in mid) or (
|
||||
|
||||
@@ -60,12 +60,15 @@ def is_embedding_model(model_id: str) -> bool:
|
||||
def is_mantle_model(model_id: str) -> bool:
|
||||
"""Pure: ids that must be probed via the mantle Responses API rather than
|
||||
Converse/InvokeModel. Matches the BARE proprietary GPT ids (`openai.gpt-5*`,
|
||||
`openai.gpt-6*` — covers GPT-5.5/5.4/5.6 and GPT-6 Astra/Sol/Luna/6.1 Sol),
|
||||
which are mantle-served. Deliberately does NOT match:
|
||||
`openai.gpt-6*` — covers GPT-5.5/5.4/5.6 and GPT-6 Astra/Sol/Luna/6.1 Sol).
|
||||
The bare id is the bedrock-mantle form for every proprietary GPT model; none
|
||||
of them has a bedrock-runtime in-region form. Deliberately does NOT match:
|
||||
- gpt-oss ids (Converse-capable on bedrock-runtime), and
|
||||
- GPT-5.6/GPT-6 CRIS profile ids (`us.`/`in.`/`global.` prefixed) — those are
|
||||
bedrock-runtime targets where Converse IS supported, so falling through to
|
||||
the standard Converse probe is the correct behavior, not an accident."""
|
||||
- CRIS profile ids (`us.`/`in.`/`global.` prefixed) for any proprietary GPT
|
||||
model, GPT-5.5/5.4 included (their model cards list `us.`/`global.` ids on
|
||||
bedrock-runtime) — those are bedrock-runtime targets where Converse IS
|
||||
supported, so falling through to the standard Converse probe is the
|
||||
correct behavior, not an accident."""
|
||||
mid = model_id.lower()
|
||||
return (mid.startswith("openai.gpt-5") or mid.startswith("openai.gpt-6")) and "oss" not in mid
|
||||
|
||||
@@ -160,8 +163,8 @@ def classify_mantle_error(status: int | None, message: str) -> dict:
|
||||
f"Mantle is in-region only; for GPT-5.6 try a bedrock-runtime CRIS "
|
||||
f"id (us./in./global. prefixed); for GPT-6 Astra use us-east-1/us-west-2 "
|
||||
f"mantle or a supported us./global. runtime CRIS path. "
|
||||
f"For GPT-5.5/5.4 switch to "
|
||||
f"a supported region or a different model."}
|
||||
f"For GPT-5.5/5.4 switch to a supported mantle region, a "
|
||||
f"us./global. bedrock-runtime CRIS id, or a different model."}
|
||||
if status == 429:
|
||||
# Reaching a token-per-minute ceiling still proves we are authorized.
|
||||
return {"ok": True, "reason": "throttled_ok",
|
||||
|
||||
@@ -295,6 +295,18 @@ def test_mantle_gpt_detection_excludes_gpt_oss():
|
||||
assert bp.is_mantle_gpt("anthropic.claude-sonnet-4-6") is False
|
||||
|
||||
|
||||
def test_mantle_gpt_detection_gpt54_55_bare_is_mantle_cris_is_runtime():
|
||||
# GPT-5.4 / GPT-5.5 model cards (read 2026-11-06) list us./global. CRIS ids on
|
||||
# bedrock-runtime. is_mantle_gpt matches only the bare (mantle) id form; the
|
||||
# prefixed runtime forms are handled by the callers' explicit CRIS regex.
|
||||
assert bp.is_mantle_gpt("openai.gpt-5.4") is True
|
||||
assert bp.is_mantle_gpt("openai.gpt-5.5") is True
|
||||
assert bp.is_mantle_gpt("us.openai.gpt-5.4") is False
|
||||
assert bp.is_mantle_gpt("global.openai.gpt-5.4") is False
|
||||
assert bp.is_mantle_gpt("us.openai.gpt-5.5") is False
|
||||
assert bp.is_mantle_gpt("global.openai.gpt-5.5") is False
|
||||
|
||||
|
||||
def test_mantle_gpt_detection_includes_bare_gpt6():
|
||||
# Finding 1: bare GPT-6 ids must be recognized by the same detector that
|
||||
# routes GPT-5.x to the curated static table / mantle path.
|
||||
|
||||
@@ -126,6 +126,18 @@ def test_mantle_model_detection_excludes_gpt_oss():
|
||||
assert is_mantle_model("anthropic.claude-sonnet-4-6") is False
|
||||
|
||||
|
||||
def test_is_mantle_model_gpt54_55_cris_ids_are_runtime_not_mantle():
|
||||
# GPT-5.4 / GPT-5.5 model cards (read 2026-11-06) list us./global. CRIS ids on
|
||||
# bedrock-runtime: the bare id needs bedrock-mantle:* actions, the prefixed id
|
||||
# needs bedrock:InvokeModel against the inference-profile ARN.
|
||||
assert is_mantle_model("openai.gpt-5.4") is True
|
||||
assert is_mantle_model("openai.gpt-5.5") is True
|
||||
assert is_mantle_model("us.openai.gpt-5.4") is False
|
||||
assert is_mantle_model("global.openai.gpt-5.4") is False
|
||||
assert is_mantle_model("us.openai.gpt-5.5") is False
|
||||
assert is_mantle_model("global.openai.gpt-5.5") is False
|
||||
|
||||
|
||||
def test_is_mantle_model_matches_bare_gpt6_ids():
|
||||
assert is_mantle_model("openai.gpt-6-astra") is True
|
||||
assert is_mantle_model("openai.gpt-6-sol") is True
|
||||
|
||||
+20
-5
@@ -197,6 +197,18 @@ def test_mantle_model_detection_excludes_gpt_oss_and_cris_forms():
|
||||
assert p.is_mantle_model("amazon.nova-lite-v1:0") is False
|
||||
|
||||
|
||||
def test_is_mantle_model_gpt54_55_cris_ids_are_runtime_not_mantle():
|
||||
# GPT-5.4 / GPT-5.5 model cards (read 2026-11-06) list us./global. CRIS ids on
|
||||
# bedrock-runtime. The ID-form rule already classifies them: bare id -> mantle
|
||||
# probe, CRIS-prefixed id -> Converse probe. Lock that in for these two models.
|
||||
assert p.is_mantle_model("openai.gpt-5.4") is True
|
||||
assert p.is_mantle_model("openai.gpt-5.5") is True
|
||||
assert p.is_mantle_model("us.openai.gpt-5.4") is False
|
||||
assert p.is_mantle_model("global.openai.gpt-5.4") is False
|
||||
assert p.is_mantle_model("us.openai.gpt-5.5") is False
|
||||
assert p.is_mantle_model("global.openai.gpt-5.5") is False
|
||||
|
||||
|
||||
def test_is_mantle_model_matches_bare_gpt6_ids():
|
||||
# Bare GPT-6 ids (Astra, Sol, Luna, 6.1 Sol) are mantle-served, same as the
|
||||
# bare GPT-5.x ids above. Their CRIS forms (us./global. prefixed) stay on the
|
||||
@@ -221,17 +233,20 @@ def test_mantle_authz_error_points_at_mantle_actions_not_invoke_model():
|
||||
|
||||
|
||||
def test_mantle_model_unavailable_remedy_is_family_split():
|
||||
# Verified 2026-08-21: mantle itself is in-region only, but GPT-5.6 now has a
|
||||
# bedrock-runtime CRIS path — so the 404 remedy must offer the CRIS form for
|
||||
# 5.6 while making clear 5.5/5.4 have no prefixed form. An earlier version of
|
||||
# this test asserted the opposite (never suggest a prefix), which matched the
|
||||
# pre-2026-08-21 docs.
|
||||
# Mantle itself is in-region only, but every proprietary GPT model has a
|
||||
# bedrock-runtime CRIS path — GPT-5.6 (us./in./global., verified 2026-08-21)
|
||||
# and GPT-5.5/5.4 (us./global., model cards read 2026-11-06). The 404 remedy
|
||||
# must offer the CRIS form for all of them. An earlier version of this test
|
||||
# asserted that 5.5/5.4 have no prefixed form, which matched the pre-2026-11-06
|
||||
# model cards.
|
||||
v = p.classify_mantle_error(404, "model not found")
|
||||
assert v["ok"] is False
|
||||
assert v["reason"] == "model_unavailable"
|
||||
assert "in-region only" in v["detail"]
|
||||
assert "CRIS" in v["detail"] and "us./in./global." in v["detail"]
|
||||
assert "GPT-5.5/5.4" in v["detail"]
|
||||
assert "us./global. bedrock-runtime CRIS id" in v["detail"]
|
||||
assert "different model" in v["detail"]
|
||||
|
||||
def test_mantle_throttle_is_ok_for_preflight():
|
||||
v = p.classify_mantle_error(429, "too many tokens per minute")
|
||||
|
||||
@@ -48,9 +48,9 @@ Inference on `bedrock-mantle` for these models is governed by **two per-model, p
|
||||
| Medium | Medium | Monitor 429s against **token** throughput, not request rate; enable prompt caching |
|
||||
| High | High | Enable prompt caching first (cached input is exempt from input TPM), then request a quota increase |
|
||||
|
||||
**The `bedrock-runtime` fallback exists only for GPT-5.6 and GPT-6 (Astra, Sol, Luna, 6.1 Sol, 6.1 Terra).** GPT-5.5 and GPT-5.4 are `bedrock-mantle` only and in-region only — for them, "switch to `bedrock-runtime`" requires moving to a different model (Bedrock-native or `gpt-oss`), a model change with its own eval cost. GPT-5.6 Sol/Terra/Luna and GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra DO have a `bedrock-runtime` path via CRIS inference profiles (GPT-5.6 `us.`/`in.`/`global.`, GPT-6 `us.`/`global.` prefixed ids; the model cards recommend runtime for new applications) — a legitimate endpoint option with its own quota family, and on Global CRIS it is also the cost-parity option. Scaling levers on the mantle path, in order:
|
||||
**Every proprietary GPT model on Bedrock has a `bedrock-runtime` path via CRIS inference profiles.** GPT-5.6 Sol/Terra/Luna (`us.`/`in.`/`global.`), GPT-6 Astra/Sol/Luna/6.1 Sol/6.1 Terra (`us.`/`global.`), and GPT-5.5 / GPT-5.4 (`us.`/`global.`; per their model cards read 2026-11-06, [GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html), [GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)) all list `bedrock-runtime` with Responses, Chat Completions and Converse; none has a runtime in-region form. The model cards recommend runtime for new applications. It is a legitimate endpoint option with its own quota family, and for GPT-5.6 / GPT-6 on Global CRIS it is also the cost-parity option (the GPT-5.5 / GPT-5.4 cards publish In-Region pricing only). A same-model endpoint switch is not a model change; only moving to a Bedrock-native model or `gpt-oss` carries the eval cost of a model change. Scaling levers on the mantle path, in order:
|
||||
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 only) — 90% off cached input and exempt from the input-TPM quota
|
||||
1. Prompt caching (GPT-5.6 and GPT-6 on mantle; GPT-5.5 / GPT-5.4 list implicit caching on `bedrock-runtime` Responses only) — 90% off cached input and exempt from the input-TPM quota
|
||||
2. Exponential backoff with a bounded retry count (`max_retries` on the OpenAI SDK)
|
||||
3. Spreading load across minutes and ramping request rate gradually rather than bursting
|
||||
4. A Service Quotas increase for that model's input/output TPM in that region
|
||||
@@ -108,8 +108,8 @@ Claude models on Mantle have an additional **output TPM cap** that differs by mo
|
||||
|
||||
- For Claude migrations at medium/high volume: the 2M output TPM cap on Claude 4.7+ is the binding constraint
|
||||
- For OpenAI proprietary GPT targets: this Claude cap does not apply. Use the GPT-5.x mantle quotas or
|
||||
Astra-specific verification above. Astra / GPT-5.6 have runtime CRIS alternatives with their own
|
||||
quotas; GPT-5.5/5.4 have no runtime path
|
||||
Astra-specific verification above. Astra / GPT-5.6 / GPT-5.5 / GPT-5.4 all have runtime CRIS alternatives
|
||||
with their own quotas (`us.`/`global.` prefixed ids; no runtime in-region form)
|
||||
- For `gpt-oss` targets: these do run on `bedrock-runtime`, so standard account TPM limits and the Converse-path mitigations apply
|
||||
- When output-heavy workloads (long JSON, tool outputs, multi-step reasoning) are detected, flag the relevant cap prominently; recommend `bedrock-runtime` for production **only** when the target model actually has a `bedrock-runtime` path
|
||||
|
||||
|
||||
@@ -42,7 +42,9 @@ Two consequences that invert the previous decision logic:
|
||||
2. **The cost of a same-model move depends on the inference option — state it conditionally.** GPT-5.x In-Region and Geo
|
||||
CRIS are priced at OpenAI's _data residency_ tier, exactly 1.10x the standard list price, so those moves cost
|
||||
~10% more. **GPT-5.6 on Global CRIS is priced at the standard list price — cost parity** — available only when
|
||||
the workload has no data-residency constraint (GPT-5.5 / GPT-5.4 have no CRIS at all). See
|
||||
the workload has no data-residency constraint. GPT-5.5 / GPT-5.4 also have `us.`/`global.` CRIS ids on
|
||||
`bedrock-runtime`, but their model cards (read 2026-11-06) publish In-Region pricing only, so quote the ~10%
|
||||
figure for them on every inference option until a Global rate is published. See
|
||||
`shared/openai-on-bedrock.md`. The non-cost case — AWS commitments, IAM/VPC/CloudTrail governance, residency,
|
||||
prompt caching, one vendor relationship — carries the in-region path; never present that path as free or neutral.
|
||||
For Astra, use its separately dated Bedrock rates and verify the source provider's rate before comparing.
|
||||
@@ -72,14 +74,15 @@ GPT-5.4, the target is **the same model on Bedrock**.
|
||||
| GPT-5.6 Sol | GPT-5.6 Sol | `openai.gpt-5.6-sol` / `us.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Terra | GPT-5.6 Terra | `openai.gpt-5.6-terra` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.6 Luna | GPT-5.6 Luna | `openai.gpt-5.6-luna` / `us.` `in.` `global.` prefixed | `strong_migrate` — same model; parity on Global CRIS, else +10% |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` (mantle only) | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.5 | GPT-5.5 | `openai.gpt-5.5` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
| GPT-5.4 | GPT-5.4 | `openai.gpt-5.4` / `us.` `global.` prefixed | `strong_migrate` — same model, ~10% over OpenAI std |
|
||||
|
||||
Then apply the **region gate** below. Record `model_change: false` and the chosen path in
|
||||
`aws-design-ai.json` → `ai_architecture.code_migration`: `migration_path: "mantle_openai_chat"` for
|
||||
Astra sources using Chat Completions, or `"mantle_openai_responses"` for a selected Mantle Responses
|
||||
path. Use `migration_path: "runtime_openai_cris"` when a GPT-6 / GPT-5.6 target is served via
|
||||
`bedrock-runtime` with a CRIS id (the model cards recommend runtime for new applications).
|
||||
sources using Chat Completions on a model whose card lists it on mantle (GPT-6, GPT-5.5, GPT-5.4), or
|
||||
`"mantle_openai_responses"` for a selected Mantle Responses path. Use `migration_path: "runtime_openai_cris"`
|
||||
when a GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 target is served via `bedrock-runtime` with a CRIS id (the model
|
||||
cards recommend runtime for new applications).
|
||||
|
||||
Report the assessment honestly: `strong_migrate` here means "low-risk, well-supported move", not "cheaper".
|
||||
|
||||
@@ -105,14 +108,16 @@ regions — offer it, stating the data-residency implication of cross-region rou
|
||||
Global routes anywhere and is the cost-parity option). Only a workload that requires strictly in-region processing
|
||||
in a region outside the mantle matrix falls through to Tier 1.
|
||||
|
||||
**GPT-5.5 / GPT-5.4:** hard gate. Mantle in-region only, no CRIS, no fallback. If the target region is not in the
|
||||
matrix:
|
||||
**GPT-5.5 / GPT-5.4:** same shape as GPT-5.6, with a narrower mantle matrix (GPT-5.5: us-east-1, us-east-2;
|
||||
GPT-5.4: us-east-1, us-east-2, us-west-2, us-gov-west-1 — per their model cards, read 2026-11-06). If the target
|
||||
region is in that matrix, either endpoint works. Otherwise the `bedrock-runtime` CRIS path (`us.openai.gpt-5.5` /
|
||||
`global.openai.gpt-5.5`, `us.openai.gpt-5.4` / `global.openai.gpt-5.4`) reaches it, with the same data-residency
|
||||
caveat as GPT-5.6 (Geo stays within the geography, Global routes anywhere). There is no runtime in-region form.
|
||||
Only a workload that requires strictly in-region processing in a region outside the mantle matrix falls through:
|
||||
|
||||
1. Prefer offering a **region change** to a supported region, if no data-residency constraint pins the workload.
|
||||
This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. If the region is fixed, also offer **upgrading within the vendor to GPT-5.6** (which reaches the region via
|
||||
CRIS) alongside the cross-family option — a generation change, but same vendor and prompt idioms.
|
||||
3. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
1. Prefer offering a **region change** to a supported mantle region, if no data-residency constraint pins the
|
||||
workload. This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
|
||||
2. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.
|
||||
|
||||
Record the outcome in `regional_warnings[]`.
|
||||
|
||||
@@ -235,10 +240,11 @@ model on price.
|
||||
|
||||
### Open-weight gpt-oss
|
||||
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and, unlike the
|
||||
proprietary GPT models, **do** support `bedrock-runtime` / Converse. They sit a capability class below the GPT-5.x
|
||||
frontier tier. Recommend them when the user wants OpenAI-architecture models on the Bedrock-native runtime surface
|
||||
(for Guardrails, invocation logging, or Converse-based tooling) rather than through mantle.
|
||||
`openai.gpt-oss-120b` (0.15 / 0.60) and `openai.gpt-oss-20b` (0.07 / 0.30) remain available and support
|
||||
`bedrock-runtime` / Converse with a bare model ID (no CRIS profile needed) and `Invoke`, which the proprietary GPT
|
||||
models do not. They sit a capability class below the GPT-5.x frontier tier. Recommend them when the user wants
|
||||
OpenAI-architecture models on the Bedrock-native runtime surface (for Guardrails, invocation logging, or
|
||||
Converse-based tooling) at the lowest price rather than through mantle.
|
||||
|
||||
---
|
||||
|
||||
@@ -284,7 +290,7 @@ rather than dismissing it. For a long-context (>272K) workload the gap is much w
|
||||
| Chat Completions | Preserve supported Chat when selected | Reshape only for a selected API change; verify the model and endpoint |
|
||||
| Function calling | Supported on GPT-5.x via mantle; Claude tools via Converse | Same-model keeps tool semantics identical |
|
||||
| Reasoning effort | `reasoning={"effort": ...}` on mantle | `none` / `low` / `medium` / `high` / `xhigh` / `max` |
|
||||
| Prompt caching | GPT-5.6 only (90% off cached input) | Not listed for GPT-5.5 / 5.4; Claude has its own caching |
|
||||
| Prompt caching | GPT-5.6 (90% off cached input); GPT-5.5 / 5.4 implicit caching on `bedrock-runtime` Responses only | GPT-5.5 / 5.4 cards (read 2026-11-06) list a cache-read rate and implicit caching on runtime, none on mantle; Claude has its own caching |
|
||||
| Streaming | Supported | Verify per surface |
|
||||
| Vision | GPT-5.x (image input) or Claude / Llama 4 | Same-model preserves behavior |
|
||||
| JSON mode | Claude (excellent), Nova Pro (good) | Most models via prompt |
|
||||
@@ -339,11 +345,12 @@ and closed to new customers as of July 30, 2026.
|
||||
### When to prefer a Bedrock-native model over the mantle GPT path
|
||||
|
||||
Choose Claude / Nova via Converse when the workload needs Knowledge Bases, intelligent prompt routing, structured
|
||||
outputs, or application inference profiles — none of which the GPT path offers. Note what no longer forces a family
|
||||
switch for GPT-5.6: its `bedrock-runtime` CRIS path provides cross-region inference, invocation logs, and
|
||||
Guardrails (Converse API only; prompt caching is Responses-only on runtime). For GPT-5.5 / GPT-5.4, Guardrails,
|
||||
logging, and cross-region reach still require a different model. The mantle GPT path is the smallest change; the
|
||||
Bedrock-native path remains the most feature-complete.
|
||||
outputs, or application inference profiles — none of which the mantle GPT path offers. Note what no longer forces
|
||||
a family switch for GPT-5.6, GPT-5.5 and GPT-5.4: their `bedrock-runtime` CRIS paths provide cross-region
|
||||
inference, invocation logs, Guardrails and application inference profiles (both Converse API only), and
|
||||
structured outputs and implicit prompt caching (both Responses API only) — per the GPT-5.4 / GPT-5.5 model cards
|
||||
read 2026-11-06. Knowledge Bases and intelligent prompt routing still require a Bedrock-native model. The mantle
|
||||
GPT path is the smallest change; the Bedrock-native path remains the most feature-complete.
|
||||
|
||||
### High-spend tiering
|
||||
|
||||
@@ -364,8 +371,8 @@ alternative, not against the source provider.
|
||||
cheaper and exempt from the input-TPM quota, which relieves the binding constraint on mantle.
|
||||
|
||||
**Very high (>100M tokens/day):** tiering plus caching. Check TPM headroom per model per region early; there is no
|
||||
RPM quota and no `bedrock-runtime` fallback for GPT models, so quota increases are the scaling lever. See
|
||||
`shared/ai-migration-guardrails.md`.
|
||||
RPM quota, and the `bedrock-runtime` CRIS path for GPT-5.6 / 5.5 / 5.4 / GPT-6 carries its own quotas rather than
|
||||
adding headroom to mantle, so quota increases are the scaling lever. See `shared/ai-migration-guardrails.md`.
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -133,8 +133,8 @@ See `vendored/ai/ai-model-lifecycle.md` for lifecycle details. **Do not recommen
|
||||
| GPT-5.6 Sol | openai.gpt-5.6-sol | OpenAI | 4.40 | 22.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 8.80/33.00) |
|
||||
| GPT-5.6 Terra | openai.gpt-5.6-terra | OpenAI | 2.20 | 13.20 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 4.40/19.80) |
|
||||
| GPT-5.6 Luna | openai.gpt-5.6-luna | OpenAI | 0.22 | 1.32 | 272K | fast | active (mantle in-region + runtime CRIS; 1M tier 0.44/1.98) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle only; no 1M tier) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle only; no 1M tier) |
|
||||
| GPT-5.5 | openai.gpt-5.5 | OpenAI | 5.50 | 33.00 | 272K | frontier | active (mantle in-region + runtime CRIS; 1M tier 11.00/49.50; card read 2026-11-06) |
|
||||
| GPT-5.4 | openai.gpt-5.4 | OpenAI | 2.75 | 16.50 | 272K | flagship | active (mantle in-region + runtime CRIS; 1M tier 5.50/24.75; card read 2026-11-06) |
|
||||
| Gemma 4 31B | google.gemma-4-31b | Google | 0.14 | 0.40 | 256K | mid | active (Mantle only) |
|
||||
| Gemma 4 26B-A4B | google.gemma-4-26b-a4b | Google | 0.13 | 0.40 | 256K | efficient | active (Mantle only) |
|
||||
| Gemma 4 E2B | google.gemma-4-e2b | Google | 0.04 | 0.08 | 128K | budget | active (Mantle only) |
|
||||
@@ -353,17 +353,24 @@ Regions (mantle): us-east-1 / us-east-2 for all five; us-west-2 additionally car
|
||||
| GPT-5.5 | 5.50 | 33.00 | — | 0.55 |
|
||||
| GPT-5.4 | 2.75 | 16.50 | — | 0.275 |
|
||||
|
||||
#### Long context (1M) — GPT-5.6 only, 2.0x input / 1.5x output
|
||||
#### Long context (>272K input tokens) — GPT-5.6 at 2.0x input / 1.5x output; GPT-5.5 / GPT-5.4 per their model cards
|
||||
|
||||
| Model | Input $/1M | Output $/1M |
|
||||
| ------------- | ---------- | ----------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 |
|
||||
| Model | Input $/1M | Output $/1M | Cache read $/1M |
|
||||
| ------------- | ---------- | ----------- | --------------- |
|
||||
| GPT-5.6 Sol | 8.80 | 33.00 | — |
|
||||
| GPT-5.6 Terra | 4.40 | 19.80 | — |
|
||||
| GPT-5.6 Luna | 0.44 | 1.98 | — |
|
||||
| GPT-5.5 | 11.00 | 49.50 | 1.10 |
|
||||
| GPT-5.4 | 5.50 | 24.75 | 0.55 |
|
||||
|
||||
A workload above 272K context must be priced at the long-context tier. GPT-5.5 and GPT-5.4 have no such tier, so
|
||||
their usable window is 272K. Global CRIS is published at OpenAI's standard list price (cost parity, GPT-5.6 only —
|
||||
see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
A workload above 272K input tokens must be priced at the long-context tier, applied to all input and output tokens
|
||||
of the request. GPT-5.5 and GPT-5.4 rows read from their model cards on 2026-11-06
|
||||
([GPT-5.4](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html),
|
||||
[GPT-5.5](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html)); both cards state a
|
||||
1,050,000-token context window (128,000 max output) and list `us.`/`global.` CRIS ids on `bedrock-runtime`, but
|
||||
publish the In-Region inference option only. Global CRIS is published at OpenAI's standard list price (cost parity,
|
||||
GPT-5.6 only — see `openai-on-bedrock.md`). GovCloud (US-West) GPT-5.4 short-context is 3.375 / 20.25 (cache read
|
||||
0.3375) per the 2026-11-06 card; an earlier revision recorded 3.30 / 19.80.
|
||||
|
||||
> **One source conflict.** These models are absent
|
||||
> from the AWS Price List API entirely: querying `AmazonBedrock` returns only `gpt-oss` and GPT OSS Safeguard, and
|
||||
@@ -373,7 +380,8 @@ see `openai-on-bedrock.md`). GovCloud GPT-5.4 is 3.30 / 19.80.
|
||||
|
||||
**Prompt caching (GPT-5.6 only):** cached input read at a 90% discount, cache write at 1.25x uncached input, minimum
|
||||
1,024-token prefix, up to 4 breakpoints, ≥30-minute retention. Cached input is exempt from the input-TPM quota.
|
||||
GPT-5.5 and GPT-5.4 do not list prompt caching support.
|
||||
GPT-5.5 and GPT-5.4 list only implicit prompt caching on `bedrock-runtime` Responses (automatic, no request
|
||||
parameters; cache-read rate above, no cache-write rate) and none on mantle (model cards read 2026-11-06).
|
||||
|
||||
#### Open-weight gpt-oss
|
||||
|
||||
|
||||
@@ -3,8 +3,8 @@
|
||||
"provider": "openai",
|
||||
"verified_at": "2026-09-24",
|
||||
"verified_region": "us-east-2",
|
||||
"source": "Opus 5.5 cross-family target: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5-5.html (retrieved 2026-09-24). OpenAI and other unchanged entries retain their evidence from the superseded 2026-08-21 catalog (removed; see git history).",
|
||||
"disclaimer": "The Opus cross-family target is updated to Opus 5.5; same-model OpenAI paths and their earlier evidence remain unchanged. This is a migration target catalog, not an account-availability claim. Verify the chosen model, API, region and CRIS profile before implementation. Opus 5.5 uses CRIS on runtime outside eu-west-2, has 1M context/128K output, always-on adaptive thinking, and no Batch tier. Fable and Mythos remain excluded from default selection. Opus 5.5 is the only Opus model in the automatic selection pool; older Opus snapshots and pricing records are historical references. The runtime in-region exception in eu-west-2 was verified 2026-10-03; the original catalog and other model evidence dates remain unchanged.",
|
||||
"source": "Opus 5.5 cross-family target: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5-5.html (retrieved 2026-09-24). GPT-5.5 and GPT-5.4 rows re-read from their public model cards on 2026-11-06 (https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html, https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html): both now list bedrock-runtime (Responses, Chat Completions, Converse via us./global. CRIS ids) and Chat Completions on bedrock-mantle. Other OpenAI entries retain their evidence from the superseded 2026-08-21 catalog (removed; see git history).",
|
||||
"disclaimer": "The Opus cross-family target is updated to Opus 5.5; the GPT-5.5 and GPT-5.4 entries were re-read from their public model cards on 2026-11-06 (runtime CRIS paths, Chat Completions on mantle, 1.05M context, long-context pricing); the other same-model OpenAI paths and their earlier evidence remain unchanged. This is a migration target catalog, not an account-availability claim. Verify the chosen model, API, region and CRIS profile before implementation. Opus 5.5 uses CRIS on runtime outside eu-west-2, has 1M context/128K output, always-on adaptive thinking, and no Batch tier. Fable and Mythos remain excluded from default selection. Opus 5.5 is the only Opus model in the automatic selection pool; older Opus snapshots and pricing records are historical references. The runtime in-region exception in eu-west-2 was verified 2026-10-03; the original catalog and other model evidence dates remain unchanged.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"sdk": "openai",
|
||||
@@ -19,8 +19,8 @@
|
||||
"GPT-5.x on Bedrock Mantle is served through the Responses API"
|
||||
],
|
||||
"constraints": [
|
||||
"Mantle itself has no cross-region form for any GPT model (in-region only); for GPT-5.6 the bedrock-runtime CRIS path covers additional regions instead (see runtime_converse)",
|
||||
"GPT-5.x on Mantle rejects Chat Completions; a Chat Completions source must reshape to Responses (reference-verified on gpt-5.4; the 5.6 cards list Chat Completions as a supported API but its mantle-side availability is unprobed — verify in the target account)",
|
||||
"Mantle itself has no cross-region form for any GPT model (in-region only); for GPT-5.6, GPT-5.5 and GPT-5.4 the bedrock-runtime CRIS path covers additional regions instead (see runtime_converse)",
|
||||
"Chat Completions support on Mantle is model-specific: the GPT-5.4 and GPT-5.5 model cards list Chat Completions as supported on bedrock-mantle (read 2026-11-06); the GPT-5.6 Sol card does not. A Chat Completions source reshapes to Responses only when the selected model's card does not list Chat Completions — verify in the target account",
|
||||
"Availability must be probed in the target account and region"
|
||||
]
|
||||
},
|
||||
@@ -36,7 +36,7 @@
|
||||
"OpenAI-compatible Chat Completions surface for models that accept it"
|
||||
],
|
||||
"constraints": [
|
||||
"Not available for the GPT-5.x models listed here: rejected on gpt-5.4 (internal reference probe) and unsupported on the gpt-5.6-sol model card",
|
||||
"Model-specific: listed as supported on bedrock-mantle by the GPT-5.4 and GPT-5.5 model cards (https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html, https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html, read 2026-11-06) and by the GPT-6 cards; unsupported on the gpt-5.6-sol model card. Uses /openai/v1/chat/completions",
|
||||
"Availability and model support must be probed in the target account and region"
|
||||
]
|
||||
},
|
||||
@@ -51,10 +51,10 @@
|
||||
"strengths": [
|
||||
"Bedrock-native multi-model request surface",
|
||||
"Supports Bedrock Guardrails (Converse API only), invocation logging, and CloudWatch",
|
||||
"For GPT-5.6 this is a SAME-MODEL governance path via CRIS (us./in./global. ids; base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs); the model cards recommend runtime for new applications, and Global CRIS is priced at OpenAI's standard rate (cost parity)"
|
||||
"For GPT-5.6, GPT-5.5 and GPT-5.4 this is a SAME-MODEL governance path via CRIS (us./in./global. ids; base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs); the model cards recommend runtime for new applications, and for GPT-5.6 Global CRIS is priced at OpenAI's standard rate (cost parity)"
|
||||
],
|
||||
"constraints": [
|
||||
"GPT-5.5 / GPT-5.4 have NO runtime path — a Converse target for them means a Bedrock-native model and a boto3 rewrite of the OpenAI integration",
|
||||
"GPT-5.5 / GPT-5.4 on runtime are CRIS-only (us. and global. ids; no in-region form) per their model cards (read 2026-11-06); the cards publish In-Region (1.10x) pricing only, so do not assume Global CRIS price parity for them",
|
||||
"GPT-5.6 on runtime is CRIS-only (no in-region form); Geo CRIS stays within the geography, Global CRIS routes anywhere — pick by data-residency posture. Runtime does not support server-side tool use, structured outputs, or application inference profiles for these models; prompt caching is Responses-only",
|
||||
"Bedrock-native models generally require a Global or geography-scoped CRIS inference profile",
|
||||
"Availability must be probed in the target account and region"
|
||||
@@ -111,19 +111,47 @@
|
||||
"generation": "reasoning",
|
||||
"version": "5.5",
|
||||
"class": "reasoning",
|
||||
"context_window": 272000,
|
||||
"output_token_ceiling": "unknown",
|
||||
"context_window": 1050000,
|
||||
"output_token_ceiling": 128000,
|
||||
"inference_profiles": {
|
||||
"us": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1"],
|
||||
"global": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1", "mx-central-1", "eu-central-1", "eu-central-2", "eu-north-1", "eu-south-1", "eu-south-2", "eu-west-1", "eu-west-2", "eu-west-3", "ap-east-2", "ap-northeast-1", "ap-northeast-2", "ap-northeast-3", "ap-south-1", "ap-south-2", "ap-southeast-1", "ap-southeast-2", "ap-southeast-3", "ap-southeast-4", "ap-southeast-5", "ap-southeast-6", "ap-southeast-7", "il-central-1", "me-central-1", "me-south-1", "af-south-1", "sa-east-1"]
|
||||
},
|
||||
"region_availability": {
|
||||
"commercial": ["us-east-1", "us-east-2"],
|
||||
"govcloud": [],
|
||||
"evidence": "bedrock-mantle in-region rows of the GPT-5.5 model card, read 2026-11-06: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html"
|
||||
},
|
||||
"pricing": {
|
||||
"source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html (read 2026-11-06); USD per 1M tokens, Standard tier; the card lists the In-Region inference option only",
|
||||
"commercial_short_context_le_272k": {"input": 5.50, "cache_read": 0.55, "output": 33.00},
|
||||
"commercial_long_context_gt_272k": {"input": 11.00, "cache_read": 1.10, "output": 49.50},
|
||||
"service_tiers": {"standard": true, "priority": false, "flex": false, "reserved": false}
|
||||
},
|
||||
"capabilities": [
|
||||
"reasoning",
|
||||
"tool_or_function_calling"
|
||||
"tool_or_function_calling",
|
||||
"image_input_vision"
|
||||
],
|
||||
"capability_source": "AWS model card 2026-07-21 (launch 2026-06-01, Active): advanced coding/agentic tasks; tool calling via the Responses API's built-in tool use. Feature-level Mantle behavior not probed — verify before relying. Re-verified 2026-08-21: still mantle-only, in-region only; no CRIS, no runtime path, no long-context tier.",
|
||||
"capability_source": "AWS model card https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-55.html (read 2026-11-06; launch 2026-06-01, Active, EOL no sooner than 2027-06-01): text+image input, text output; 1,050,000-token context shared between input and output, 128,000 max output. Mantle lists server-side and client-side tool calling; runtime lists invocation logs, streaming, Guardrails (Converse only), application inference profiles (Converse only), implicit prompt caching (Responses only), structured outputs (Responses) and reasoning summaries; server-side tool use is unsupported on runtime. Documentation claims, not account probes — verify before relying.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": false,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (2026-07-21): bedrock-mantle Programmatic Access with Responses sample code; Geo/Global inference not supported"
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): bedrock-mantle supports Responses at /openai/v1/responses; in-region us-east-1 and us-east-2 only; Geo/Global not supported on mantle"
|
||||
},
|
||||
"mantle_openai_chat": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": false,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): Chat Completions listed as supported on bedrock-mantle at /openai/v1/chat/completions, with sample code"
|
||||
},
|
||||
"runtime_converse": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.5",
|
||||
"requires_cris": true,
|
||||
"evidence": "AWS model card model-card-openai-gpt-55 (read 2026-11-06): bedrock-runtime supported with Responses, Chat Completions and Converse (Messages and Invoke unsupported); no in-region form — requests must name us.openai.gpt-5.5 (Geo) or global.openai.gpt-5.5 (Global); base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs"
|
||||
}
|
||||
}
|
||||
},
|
||||
@@ -213,27 +241,50 @@
|
||||
"generation": "reasoning",
|
||||
"version": "5.4",
|
||||
"class": "reasoning",
|
||||
"context_window": "unknown",
|
||||
"output_token_ceiling": "unknown",
|
||||
"context_window": 1050000,
|
||||
"output_token_ceiling": 128000,
|
||||
"inference_profiles": {
|
||||
"us": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1"],
|
||||
"global": ["us-east-1", "us-east-2", "us-west-1", "us-west-2", "ca-central-1", "ca-west-1", "mx-central-1", "eu-central-1", "eu-central-2", "eu-north-1", "eu-south-1", "eu-south-2", "eu-west-1", "eu-west-2", "eu-west-3", "ap-east-2", "ap-northeast-1", "ap-northeast-2", "ap-northeast-3", "ap-south-1", "ap-south-2", "ap-southeast-1", "ap-southeast-2", "ap-southeast-3", "ap-southeast-4", "ap-southeast-5", "ap-southeast-6", "ap-southeast-7", "il-central-1", "me-central-1", "me-south-1", "af-south-1", "sa-east-1"]
|
||||
},
|
||||
"region_availability": {
|
||||
"commercial": ["us-east-1", "us-east-2", "us-west-2"],
|
||||
"govcloud": ["us-gov-west-1"],
|
||||
"evidence": "bedrock-mantle in-region rows of the GPT-5.4 model card, read 2026-11-06: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html"
|
||||
},
|
||||
"pricing": {
|
||||
"source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html (read 2026-11-06); USD per 1M tokens, Standard tier; the card lists the In-Region inference option only",
|
||||
"commercial_short_context_le_272k": {"input": 2.75, "cache_read": 0.275, "output": 16.50},
|
||||
"commercial_long_context_gt_272k": {"input": 5.50, "cache_read": 0.55, "output": 24.75},
|
||||
"govcloud_us_west_short_context_le_272k": {"input": 3.375, "cache_read": 0.3375, "output": 20.25},
|
||||
"service_tiers": {"standard": true, "priority": false, "flex": false, "reserved": false}
|
||||
},
|
||||
"capabilities": [
|
||||
"tool_or_function_calling",
|
||||
"structured_output_json",
|
||||
"reasoning",
|
||||
"streaming"
|
||||
"streaming",
|
||||
"image_input_vision"
|
||||
],
|
||||
"capability_source": "internal reference smoke tests and its reasoning-API migration notebook (Responses path, 2026-06 probes) — the only entry with feature-level Mantle behavior probes (tools, structured output, sampling); limits not stated Re-verified 2026-08-21: still mantle-only, in-region only; no CRIS, no runtime path, no long-context tier.",
|
||||
"capability_source": "AWS model card https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html (read 2026-11-06; launch 2026-06-01, Active, EOL no sooner than 2027-06-01): text+image input, text output; 1,050,000-token context shared between input and output, 128,000 max output. Mantle lists server-side and client-side tool calling; runtime lists invocation logs, response streaming, Guardrails (Converse only), application inference profiles (Converse only), implicit prompt caching (Responses only), structured outputs (JSON Schema; Responses API) and reasoning summaries; server-side tool use is unsupported on runtime. Documentation claims, not account probes — verify before relying.",
|
||||
"paths": {
|
||||
"mantle_openai_responses": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": false,
|
||||
"evidence": "internal reference smoke-test default Mantle model id; client.responses.create verified live"
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): bedrock-mantle supports Responses at /openai/v1/responses; in-region us-east-1, us-east-2, us-west-2 and us-gov-west-1; Geo/Global not supported on mantle"
|
||||
},
|
||||
"mantle_openai_chat": {
|
||||
"available": false,
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": false,
|
||||
"evidence": "internal reference verified gotcha: GPT-5.x on Mantle is Responses-only; Chat Completions rejected"
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): Chat Completions listed as supported on bedrock-mantle at /openai/v1/chat/completions, with sample code"
|
||||
},
|
||||
"runtime_converse": {
|
||||
"available": true,
|
||||
"model_id": "openai.gpt-5.4",
|
||||
"requires_cris": true,
|
||||
"evidence": "AWS model card model-card-openai-gpt-54 (read 2026-11-06): bedrock-runtime supported with Responses, Chat Completions and Converse (Messages and Invoke unsupported); no in-region form — requests must name us.openai.gpt-5.4 (Geo) or global.openai.gpt-5.4 (Global); base URL bedrock-runtime.{region}.amazonaws.com/openai/v1 for the OpenAI-compatible APIs"
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
@@ -190,12 +190,18 @@ Traditional-AI capabilities (AWS AI service, not Bedrock): `document_extraction`
|
||||
## migration_path vocabulary
|
||||
|
||||
`mantle_openai_responses` (an OpenAI-protocol source — direct or Azure-fronted — with the model
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `converse`
|
||||
(Bedrock-native Converse/`bedrock-runtime`; set `model_change: true` when a proprietary GPT
|
||||
model must move off mantle for Guardrails / Knowledge Bases / logging / an unsupported region) ·
|
||||
`gpt-oss` (OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic
|
||||
Bedrock SDK swap). A proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — those models are mantle-only.
|
||||
on Bedrock and the region carrying it: keep the SDK, `model_change: false`) · `mantle_openai_chat`
|
||||
(same, preserving a Chat Completions source on a model whose card lists Chat Completions on
|
||||
mantle — GPT-6, GPT-5.5, GPT-5.4) · `runtime_openai_cris` (same-model move served via
|
||||
`bedrock-runtime` with a `us.`/`in.`/`global.` CRIS id, `model_change: false`; available for
|
||||
GPT-6, GPT-5.6, GPT-5.5 and GPT-5.4 per their model cards) · `converse` (Bedrock-native
|
||||
Converse/`bedrock-runtime`; set `model_change: true` when the workload needs Knowledge Bases,
|
||||
intelligent prompt routing, or another feature no same-model path serves) · `gpt-oss`
|
||||
(OpenAI-lineage model on the Bedrock-native runtime) · `direct` (framework-agnostic Bedrock SDK
|
||||
swap). A **bare** proprietary `openai.gpt-*` model ID is NEVER paired with a
|
||||
`converse`/`bedrock-runtime` path — bare ids are mantle ids; every proprietary GPT target on
|
||||
`bedrock-runtime` carries a CRIS-prefixed id (there is no runtime in-region form for GPT-6,
|
||||
GPT-5.6, GPT-5.5 or GPT-5.4).
|
||||
|
||||
## regional_validation
|
||||
|
||||
@@ -255,8 +261,9 @@ rubric file (`ai.md`, a provider-to-Bedrock mapping guide, or a `vendored/ai/*`
|
||||
`price_comparison`.
|
||||
- [ ] `honest_assessment` overall is the weakest across `bedrock_models[]`; if
|
||||
`recommend_stay`, `honest_assessment_reason` names a non-cost blocker.
|
||||
- [ ] No `bedrock_models[]` entry pairs a proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (mantle-only).
|
||||
- [ ] No `bedrock_models[]` entry pairs a **bare** proprietary `openai.gpt-*` model ID with a
|
||||
`converse`/`bedrock-runtime` migration path (bare ids are mantle ids; a `bedrock-runtime`
|
||||
target for GPT-6 / GPT-5.6 / GPT-5.5 / GPT-5.4 carries a `us.`/`in.`/`global.` CRIS id).
|
||||
- [ ] For a retained source whose `source_service` is `azure_openai`, or whose provider resolves
|
||||
to direct OpenAI: every source model that is available on Bedrock and carried by the
|
||||
target region maps to ITSELF with `model_change: false` — not to a Claude/Nova substitute.
|
||||
|
||||
@@ -70,15 +70,25 @@ def test_both_estimate_ai_files_require_matching_context_tiers() -> None:
|
||||
|
||||
# --- Finding 2: runtime_openai_cris gate/dispatch ---
|
||||
|
||||
def test_design_ai_scopes_mantle_only_to_gpt55_gpt54() -> None:
|
||||
def test_design_ai_gives_gpt55_gpt54_the_runtime_cris_path() -> None:
|
||||
# GPT-5.5 / GPT-5.4 model cards (read 2026-11-06) list us./global. CRIS ids on
|
||||
# bedrock-runtime, so the runtime_openai_cris carve-out must name them alongside
|
||||
# GPT-6/5.6 and the old "mantle-only, no CRIS" scoping must be gone.
|
||||
for f in DESIGN_AI_FILES:
|
||||
text = f.read_text(encoding="utf-8")
|
||||
text = re.sub(r"\s+", " ", f.read_text(encoding="utf-8"))
|
||||
assert "runtime_openai_cris" in text, (
|
||||
f"{f}: no runtime_openai_cris carve-out for GPT-6/5.6 (Finding 2)"
|
||||
f"{f}: no runtime_openai_cris carve-out (Finding 2)"
|
||||
)
|
||||
assert "GPT-5.5" in text and "GPT-5.4" in text, (
|
||||
f"{f}: mantle-only scoping must name GPT-5.5/GPT-5.4 specifically (Finding 2)"
|
||||
f"{f}: the runtime_openai_cris rule must name GPT-5.5/GPT-5.4 specifically"
|
||||
)
|
||||
assert "us.openai.gpt-5.5" in text and "us.openai.gpt-5.4" in text, (
|
||||
f"{f}: must record the GPT-5.5/GPT-5.4 us. CRIS ids"
|
||||
)
|
||||
assert "GPT-5.5 and GPT-5.4 are mantle-only" not in text
|
||||
assert "GPT-5.5 and GPT-5.4 are `bedrock-mantle` only" not in text
|
||||
assert "no CRIS" not in text
|
||||
assert "No Converse fallback for GPT-5.5/GPT-5.4" not in text
|
||||
|
||||
|
||||
def test_generate_artifacts_ai_dispatches_runtime_openai_cris() -> None:
|
||||
@@ -102,9 +112,16 @@ def test_schema_design_aws_ai_vocabulary_includes_runtime_openai_cris() -> None:
|
||||
"schema-design-aws-ai.md: migration_path vocabulary missing runtime_openai_cris (Finding 2)"
|
||||
)
|
||||
assert "GPT-5.5" in text and "GPT-5.4" in text, (
|
||||
"schema-design-aws-ai.md: mantle-only scoping must name GPT-5.5/GPT-5.4 specifically, "
|
||||
"not blanket 'proprietary GPT' (Finding 2)"
|
||||
"schema-design-aws-ai.md: runtime_openai_cris vocabulary must name GPT-5.5/GPT-5.4 "
|
||||
"specifically, not blanket 'proprietary GPT' (Finding 2)"
|
||||
)
|
||||
# The ban is on BARE proprietary GPT ids paired with converse/bedrock-runtime,
|
||||
# not on the GPT-5.5/GPT-5.4 models — their cards list us./global. CRIS ids.
|
||||
normalized = re.sub(r"\s+", " ", text)
|
||||
assert "those models are mantle-only" not in normalized
|
||||
assert "mantle-only, no CRIS" not in normalized
|
||||
assert "**bare** proprietary `openai.gpt-*`" in normalized
|
||||
assert "us.openai.gpt-5.5" in normalized or "GPT-5.5 / GPT-5.4 runtime support" in normalized
|
||||
|
||||
|
||||
# --- Finding 3: Astra Invoke-support correction ---
|
||||
|
||||
@@ -114,6 +114,12 @@ def test_behavior_delta_no_longer_claims_chat_completions_unverified() -> None:
|
||||
assert "OPTIONAL" in text, (
|
||||
"openai-to-bedrock.md must make the Responses reshape optional, not mandatory"
|
||||
)
|
||||
# GPT-5.5 / GPT-5.4 model cards (read 2026-11-06) list Chat Completions on both
|
||||
# bedrock-mantle and bedrock-runtime — the reshape is optional for them too.
|
||||
assert "GPT-5.5 / GPT-5.4 follow the same rule" in text
|
||||
assert "model-card-openai-gpt-54.html" in text and "model-card-openai-gpt-55.html" in text
|
||||
assert "mantle-only, no CRIS" not in text
|
||||
assert "still unconfirmed" not in text
|
||||
|
||||
|
||||
def test_both_design_ai_no_longer_force_gpt6_chat_completions_reshape() -> None:
|
||||
@@ -122,12 +128,17 @@ def test_both_design_ai_no_longer_force_gpt6_chat_completions_reshape() -> None:
|
||||
assert "near-drop-in" in text.lower() or "optional" in text.lower(), (
|
||||
f"{f}: a GPT-6/GPT-5.6 Chat Completions source must not be forced to reshape"
|
||||
)
|
||||
assert "GPT-5.5" in text and "GPT-5.4" in text, (
|
||||
f"{f}: the mandatory-reshape exception must still be scoped to GPT-5.5/GPT-5.4"
|
||||
assert "GPT-5.5 and GPT-5.4 follow the same rule" in text, (
|
||||
f"{f}: GPT-5.5/GPT-5.4 Chat Completions is listed on both endpoints (cards read "
|
||||
"2026-11-06); they must not be the mandatory-reshape exception"
|
||||
)
|
||||
assert "GPT-5.5 and GPT-5.4 are the exception" not in text
|
||||
assert "unconfirmed" not in text
|
||||
|
||||
|
||||
def test_gcp_generate_ai_no_longer_forces_gpt6_chat_completions_reshape() -> None:
|
||||
text = _norm(GCP_GENERATE_AI_MD.read_text(encoding="utf-8"))
|
||||
assert "optional" in text.lower()
|
||||
assert "GPT-5.5" in text and "GPT-5.4" in text
|
||||
assert "GPT-5.5/GPT-5.4 follow the same rule" in text
|
||||
assert "still need a reshape" not in text
|
||||
assert "unconfirmed" not in text
|
||||
|
||||
@@ -2,8 +2,9 @@
|
||||
GPT-6 recognition, report pricing bugs).
|
||||
|
||||
Finding 0 — runtime_openai_cris carve-out gaps:
|
||||
(a) Azure design.md / generate.md _assert blocks must scope the GPT-ban to
|
||||
GPT-5.5/GPT-5.4 and explicitly carve out GPT-5.6/GPT-6 on runtime_openai_cris.
|
||||
(a) Azure design.md / generate.md _assert blocks must scope the GPT-ban to BARE
|
||||
proprietary ids and explicitly carve out GPT-5.6/GPT-6 — and, since the
|
||||
2026-11-06 model-card re-read, GPT-5.5/GPT-5.4 — on runtime_openai_cris.
|
||||
(b) GCP generate-ai.md Part 1 must route runtime_openai_cris through
|
||||
bedrock-runtime, not the mantle URL/IAM.
|
||||
(c) Canonical model-selection guide + vendored guardrails must name the full
|
||||
@@ -79,21 +80,30 @@ def _norm(text: str) -> str:
|
||||
# --- Finding 0.1: Azure design.md / generate.md assert carve-out ---
|
||||
|
||||
|
||||
def test_azure_design_assert_scopes_gpt_ban_to_55_54_and_carves_out_cris() -> None:
|
||||
def test_azure_design_assert_scopes_gpt_ban_to_bare_ids_and_carves_out_cris() -> None:
|
||||
# GPT-5.5 / GPT-5.4 model cards (read 2026-11-06) list us./global. CRIS ids on
|
||||
# bedrock-runtime, so the ban is on BARE proprietary ids paired with
|
||||
# converse/bedrock-runtime — not on the GPT-5.5/GPT-5.4 models themselves.
|
||||
text = _norm(AZURE_DESIGN_MD.read_text(encoding="utf-8"))
|
||||
assert "no proprietary GPT-5.5 or GPT-5.4 model ID is paired" in text, (
|
||||
"design.md _assert must scope the ban to GPT-5.5/GPT-5.4, not blanket 'openai.gpt-*'"
|
||||
assert "no bare proprietary openai.gpt-* model ID (including bare openai.gpt-5.5 or openai.gpt-5.4) is paired" in text, (
|
||||
"design.md _assert must ban bare proprietary ids on converse/bedrock-runtime, "
|
||||
"naming bare openai.gpt-5.5/openai.gpt-5.4 explicitly"
|
||||
)
|
||||
assert "runtime_openai_cris on bedrock-runtime — this is not a violation" in text
|
||||
assert "GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation" in text
|
||||
assert "GPT-5.5/GPT-5.4 us./global. profiles" in text
|
||||
assert "no proprietary GPT-5.5 or GPT-5.4 model ID is paired" not in text
|
||||
assert "no proprietary openai.gpt-* model ID is paired" not in text
|
||||
|
||||
|
||||
def test_azure_generate_assert_scopes_gpt_ban_to_55_54_and_carves_out_cris() -> None:
|
||||
def test_azure_generate_assert_scopes_gpt_ban_to_bare_ids_and_carves_out_cris() -> None:
|
||||
text = _norm(AZURE_GENERATE_MD.read_text(encoding="utf-8"))
|
||||
assert "no proprietary GPT-5.5 or GPT-5.4 model ID is paired" in text, (
|
||||
"generate.md _assert must scope the ban to GPT-5.5/GPT-5.4, not blanket 'openai.gpt-*'"
|
||||
assert "no bare proprietary openai.gpt-* model ID (including bare openai.gpt-5.5 or openai.gpt-5.4) is paired" in text, (
|
||||
"generate.md _assert must ban bare proprietary ids on converse/bedrock-runtime, "
|
||||
"naming bare openai.gpt-5.5/openai.gpt-5.4 explicitly"
|
||||
)
|
||||
assert "runtime_openai_cris on bedrock-runtime — this is not a violation" in text
|
||||
assert "GPT-5.6, GPT-6, GPT-5.5 and GPT-5.4 models MAY use migration_path: runtime_openai_cris on bedrock-runtime — this is not a violation" in text
|
||||
assert "GPT-5.5/GPT-5.4 us./global. profiles" in text
|
||||
assert "no proprietary GPT-5.5 or GPT-5.4 model ID is paired" not in text
|
||||
assert "no proprietary openai.gpt-* model ID is paired" not in text
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user