feat(savings): add the per-request counterfactual savings record (#4236)

## Description

Part of #4238 (design + PR plan).

Step 1 of the savings redesign. This PR adds
`headroom/proxy/savings.py`, a pure module that prices one request's
removal as **counterfactual input cost − actual input cost**. It is not
wired into any surface yet. A follow-up PR makes /stats, the durable
ledger, `persistent_savings`, the dashboard and PERF all read this one
record, so no surface computes savings its own way.

### Why

The test: we ran Headroom Enterprise 0.4.3 in a Daytona sandbox with
Codex CLI 0.162.1 on `gpt-5.6-sol` and the real GitHub MCP server, doing
workshop task 1. We repeated the run with Claude Code on
`claude-sonnet-5`. Each call was audited against the provider's own
billing.

What we found:

| Run | Truth | /stats `cost.savings_usd` | `persistent_savings` |
|---|---|---|---|
| Codex | $0.0850 saved, 98,188 tokens | $0.0434 (−49%) | — |
| Claude Code | $0.1883 saved, 747,493 tokens | $0.0874 (−54%) | $0.2654
(+41%) |

The Codex /stats figure books only the conversation-novel tokens:
re-sent history is $0 and OpenAI cache writes are ignored. On Claude
Code, the /stats figure and `persistent_savings` disagree with each
other.

Four independent errors feed into this:
- PERF `tok_before` is +4% on Codex.
- PERF `tok_before` leaves out tool schemas on Claude Code.
- Re-sent history is priced at $0 on Responses, but re-booked at the
write rate on Anthropic.
- `cache_write_tokens` is ignored.

Each surface also prices on its own.

### The record

Headroom's tokenizer gives per-item counts in Headroom units for the
ORIGINAL and the SENT body (tools, system, each input item). The
provider gives exact totals. They are only ever combined through ratios:

- **Removed tokens:**
- `R_old` is removed tokens in items within the longest prefix of the
original body that is identical to the previous original body.
  - `R_new` is removed tokens in every other item.
  - Both are signed, so an item Headroom grew counts against the saving.
- **Shared prefix and bust:**
  - `P_S` is the sent tokens shared with the previous sent body.
  - `bust` is old items whose sent copy Headroom changed.
- **Ratios:**
  - `k = B/S` converts Headroom units to provider units.
  - `warm = min(1, (C/B)/(P_S/S))`.
  - `w = W/(U+W)`.
- **Cost:**
  - `p_miss = w·p_write + (1−w)·p_in`.
- `saved = k·R_old·[warm·p_read + (1−warm)·p_miss] + k·R_new·p_miss −
k·bust·warm·(p_miss − p_read)`.
- It is computed as explicit counterfactual and actual buckets, so the
two context tiers are priced independently.
- **Provider rules:**
- Cache minimums: OpenAI needs 1024 tokens; Anthropic uses the catalog's
per-model `prompt_cache_min_tokens`, e.g. 4096 on Haiku 4.5.
  - OpenAI cache reads come in 128-token blocks.
  - 5m and 1h writes are priced at their own rates.
- **Missing usage:** priced at list and marked `estimated`.
- **Never raises.**

The bust charge is capped at what the request actually missed.
Item-level `P_S` undercounts a token-level hit, and without the cap the
counterfactual buckets went negative; this is covered by a test.

Also in this PR: `pricing.counterfactual.catalog_min_cacheable_tokens`,
with the guarded catalog lookup shared through `_catalog_row`.

## Type of Change

- [x] New feature (non-breaking change that adds functionality)
- [x] Bug fix (non-breaking change that fixes an issue): review fix
`bb1275b7`, below

## Changes Made

- `headroom/proxy/savings.py`: `ItemCounts`, `ItemMeasure`,
`SavingsRecord`, `measure_items`, `compute_savings_record` (the record
described above).
- `headroom/pricing/counterfactual.py`: `catalog_min_cacheable_tokens`,
with the guarded catalog lookup shared through `_catalog_row`.
- `tests/test_savings_record.py`: pinned-catalog tests.
- Review fix `bb1275b7` (JerrettDavis): the counterfactual's cache
buckets are now built whole instead of as deltas clamped onto the actual
buckets.
- **Bug:** take a 1,000-token original item that Headroom sent as 10,000
tokens on successive requests (`ItemMeasure(10000, -9000, 0, 10000, 0)`,
`CacheMix(read=9000, uncached=1000)`, OpenAI, $3/M in and $0.30/M read).
The counterfactual inherited the actual reads and priced 1,000 tokens as
reads: **$0.0003**. That prompt is below OpenAI's 1,024-token minimum,
so it should cost **$0.003**.
- **Fix: buckets.** The counterfactual's read, write and uncached
buckets are nonnegative and sum to its billed input (`k × original`).
- **Fix: reads.** Reads are the actual reads, plus the removed or busted
old tokens that would have hit, minus the reads a grown old item gave
up. They never exceed the counterfactual's billed input, and below its
own cache minimum there are none. OpenAI reads come in 128-token blocks.
- **Fix: writes.** A counterfactual smaller than anything shown
cacheable has no writes.
- **Catalog minimum.** The catalog's `prompt_cache_min_tokens` is used
unless this request was actually cached or written below it. In that
case the catalog is wrong and no minimum is applied, which is the
direction that cannot inflate the saving.
- **Pricing.** The part of each counterfactual bucket the request was
billed for prices at billing rates; only the excess prices at estimate
rates. When billing and estimate rates differ, a shrunk bucket can no
longer drive the counterfactual below zero.

## Testing

- [x] Unit tests pass (`pytest`)
- [x] Linting passes (`ruff check .`)
- [x] Type checking passes (`mypy headroom`)
- [x] New tests added for new functionality

- [x] Unit tests pass. New file `tests/test_savings_record.py` (12
tests) uses a pinned catalog. It covers:
- the worked example: k=1.0526, warm=0.97159, w=0.8, saved=$0.058618,
matching the closed form to 1e-9 and also cf − actual;
  - a bust, an over-large bust, a cold cache, and 1h writes;
  - an inferred OpenAI write, and OpenAI 128-token blocks;
- the per-model cache minimum, and the counterfactual's own context
tier;
  - missing usage, an unpriced model, and the old/new item split.
- [x] Related suites pass: `test_counterfactual_pricing`,
`test_cost_cache_prices`, savings, cost and ledger suites (473 passed).
- [x] `ruff check` / `ruff format --check` (0.16.8) and `mypy headroom`
are clean.

Review fix `bb1275b7` adds regression tests that assert cost, not just
`warm`. **Fail before** is `bb1275b7`'s tests run against the parent's
`savings.py`: 8 failed. **Pass after** is all 29 passing.

- **Jerrett's exact case**, on OpenAI and Anthropic: actual $0.0057,
counterfactual $0.003, saved −$0.0027. Before the fix the counterfactual
was $0.0003.
- **OpenAI boundaries:** originals of 1,023 / 1,024 / 1,151 / 1,152 /
2,000 tokens give 0 / 1,024 / 1,024 / 1,152 / 1,920 reads. Before the
fix, 1,023 → 1,023 reads, 1,151 → 1,151, and 2,000 → 2,000.
- **Anthropic (Haiku, minimum 4,096):** originals of 4,095 / 4,096 /
4,097 tokens give 0 / 4,096 / 4,097 reads.
- **Too small to cache, with writes:** costs 1,000 × input with no write
premium. Before the fix it was 1,000 × write.
- **A grown old item alongside new removed items** loses its reads.
Before the fix it kept all 9,000.
- **No negative cost:** the counterfactual stays ≥ 0 when billing and
estimate read rates differ. Before the fix it was −$0.0078 in the
reviewer's probe.
- **Guards that pass on both versions:** a catalog minimum the provider
disproved is not applied; nothing removed saves exactly $0 on OpenAI and
Anthropic.
- **Unchanged:** the worked example, $0.058618.

### Test Output

```text
$ UV_FROZEN=1 pytest tests/test_savings_record.py -q
29 passed in 1.98s
# fail-before: the new tests against the parent savings.py
8 failed, 20 passed   (test_an_old_item_headroom_grew_inherits_no_reads_below_the_cache_minimum: Obtained 0.0003, expected 0.003; ...)
$ UV_FROZEN=1 pytest tests -q -k "savings or counterfactual or cost or ledger"
730 passed, 8 skipped, 16437 deselected
$ ruff check . && ruff format --check headroom tests     (ruff 0.17.0, as pinned in .pre-commit-config.yaml)
All checks passed!
$ mypy headroom --ignore-missing-imports
headroom/integrations/langchain/chat_model.py:380: error: Function "RunnableBinding" could always be true  [truthy-function]
# the one error is in a file this PR does not touch (local langchain version); savings.py and counterfactual.py are clean
```

## Real Behavior Proof

- Environment: macOS, this branch, the live litellm catalog
(`claude-sonnet-4-5`: $3/M input, $0.30/M read, $3.75/M write).
- Exact command / steps: `compute_savings_record(ItemMeasure(95_000,
50_000, 10_000, 88_000, 0), CacheMix(uncached=2_000, read=90_000,
write_5m=8_000), model="claude-sonnet-4-5", provider="anthropic")`
- Observed result: `k=1.0526 warm=0.97159 w=0.80 saved_usd=0.058618
actual=0.063000 cf=0.121618`. With an over-large bust,
`saved_usd=-0.003075`; the counterfactual stays ≥ 0.
- Not tested: real traffic and provider-billed comparison. The module is
not called by the proxy until the wiring PR, and the ±5% audit gate
(Codex $0.0850, Claude Code $0.1883) lands with the audit eval PR.

Review fix, on the live litellm catalog
(`LITELLM_LOCAL_MODEL_COST_MAP=True`, `gpt-4o`: $2.50/M input, $1.25/M
read), with Jerrett's measure:

| | actual | counterfactual | saved |
|---|---|---|---|
| before (`5db1b343`) | $0.013750 | $0.001250 | −$0.012500 |
| after (`bb1275b7`) | $0.013750 | $0.002500 | −$0.011250 |

The worked example on `claude-sonnet-4-5` is unchanged:
`saved_usd=0.058618`.

## Runtime Rollout Safety

- Rollout-managed feature(s): none
- Minimum rollout channel: n/a
- Stable/default behavior changed: no. The module is not called yet.
- Kill switch / disable path: none needed; nothing calls the module
until the wiring PR
- Unsafe override required: no
- Qualification impact: none
- Rollback path: revert.

## Review Readiness

- [x] I have performed a self-review
- [x] This PR is ready for human review

Review notes: Plus `code-reviewer` passes on the review fix; its three
BLOCKERs and two SHOULDs are fixed, and its final pass had no findings.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: JD Davis <mxjerrett@gmail.com>
This commit is contained in:
Tejas Chopra
2026-10-16 16:50:43 -05:00
committed by GitHub
co-authored by Claude Opus 5.5 JD Davis
parent 46306024e7
commit 2912b7b620
7 changed files with 1542 additions and 15 deletions
+37 -15
View File
@@ -374,6 +374,26 @@ def _canonical_cache_rate(litellm: Any, model: str, field: str, *, long_context:
return rate if math.isfinite(rate) and rate >= 0 else 0.0
def _catalog_row(model: str) -> tuple[Any, str, dict[str, Any]] | None:
"""``(litellm, resolved model, catalog row)``, or ``None``. Never raises."""
litellm = _litellm()
if litellm is None:
return None
try:
from headroom.pricing.litellm_pricing import resolve_litellm_model
resolved_model = resolve_litellm_model(model)
info = litellm.model_cost.get(resolved_model, {}) or {}
except Exception as exc:
logger.debug(
"counterfactual: LiteLLM catalog lookup failed for %s: %s",
safe_id(model),
describe_exception(exc),
)
return None
return litellm, resolved_model, info
@lru_cache(maxsize=_RATE_CACHE_MAXSIZE)
def resolve_rates(
model: str,
@@ -413,22 +433,10 @@ def resolve_rates(
catalog between cases must call ``resolve_rates.cache_clear()`` in between,
or the previous case's rates leak into the next.
"""
litellm = _litellm()
if litellm is None:
return None
try:
from headroom.pricing.litellm_pricing import resolve_litellm_model
resolved_model = resolve_litellm_model(model)
info = litellm.model_cost.get(resolved_model, {}) or {}
except Exception as exc:
logger.debug(
"counterfactual: LiteLLM catalog lookup failed for %s: %s",
safe_id(model),
describe_exception(exc),
)
row = _catalog_row(model)
if row is None:
return None
litellm, resolved_model, info = row
base = info.get("input_cost_per_token")
# `is None` distinguishes "unknown model" from "genuinely free". `not base`
@@ -530,6 +538,19 @@ def resolve_rates(
)
def catalog_min_cacheable_tokens(model: str) -> int | None:
"""The catalog's smallest prompt the provider will cache for ``model``.
LiteLLM publishes it as ``prompt_cache_min_tokens`` (Anthropic: 1024 for
Sonnet, 4096 for Haiku 4.5). ``None`` when litellm or the row lacks it.
"""
row = _catalog_row(model)
value = row[2].get("prompt_cache_min_tokens") if row is not None else None
if isinstance(value, bool) or not isinstance(value, int | float):
return None
return int(value) if math.isfinite(value) and value > 0 else None
@dataclass(frozen=True)
class PricedSavings:
"""What a set of counterfactual tokens was worth, and how sure we are.
@@ -641,6 +662,7 @@ __all__ = [
"PricedSavings",
"Region",
"TokenSplit",
"catalog_min_cacheable_tokens",
"price_savings",
"resolve_rates",
"split_tokens",
+33
View File
@@ -45,6 +45,7 @@ from headroom.proxy.provider_usage import (
usage_int,
)
from headroom.proxy.rate_limit_identity import rate_limit_identity
from headroom.proxy.savings import ItemMeasure, get_item_ledger, prefix_key, responses_items
from headroom.proxy.semantic_cache_key_policy import compute_request_cache_partition
from headroom.proxy.stage_timer import StageTimer, emit_stage_timings_log
from headroom.proxy.tool_schema_savings_policy import (
@@ -6391,6 +6392,36 @@ class OpenAIHandlerMixin:
"upstream request for server-side retrieval handling"
)
# Per-item original/sent counts against this conversation's previous
# request, in Headroom's tokenizer (headroom.proxy.savings). ORIGINAL
# is the client's bytes, before memory, compression or shaping; SENT is
# the body forwarded below. All of it runs off the event loop: it
# decodes, serializes and tokenizes the whole transcript. Skipped for
# incremental bodies (previous_response_id / conversation): the
# provider bills context they do not carry. It covers the first
# upstream call only; continuations and re-drives are the wiring's to
# mark as estimated. The outcome funnel settles it (ItemLedger.settle).
savings_measure = None
if not (body.get("previous_response_id") or body.get("conversation")):
def _measure_items() -> ItemMeasure:
original = responses_items(json.loads(original_body_bytes))
return get_item_ledger().measure(
prefix_key(_pre_compression_conversation_key, model, original),
original,
responses_items(body),
tokenizer.count_text,
)
try:
savings_measure = await asyncio.to_thread(_measure_items)
except Exception as exc: # noqa: BLE001 - telemetry must never break a request
logger.warning(
"[%s] savings item measurement skipped: %s",
request_id,
describe_exception(exc),
)
try:
if stream and not buffered_stream_ccr:
# Streaming for Responses API uses semantic events
@@ -6415,6 +6446,7 @@ class OpenAIHandlerMixin:
waste_signals=waste_signals_dict,
conversation_key=responses_conversation_key,
conversation_tokens_saved=tokens_saved,
savings_measure=savings_measure,
)
else:
@@ -6883,6 +6915,7 @@ class OpenAIHandlerMixin:
tags=_resp_log_tags,
conversation_key=responses_conversation_key,
conversation_tokens_saved=tokens_saved,
savings_measure=savings_measure,
turn_id=compute_turn_id(model, body.get("instructions"), messages),
request_messages=messages
if getattr(self.config, "log_full_messages", False)
+9
View File
@@ -32,6 +32,8 @@ from headroom.proxy.token_counting import gemini_output_tokens
if TYPE_CHECKING:
from fastapi.responses import Response, StreamingResponse
from headroom.proxy.savings import ItemMeasure
import httpx
@@ -815,6 +817,7 @@ class StreamingMixin:
# See ``conversation_savings``.
conversation_key: str | None = None,
conversation_tokens_saved: int | None = None,
savings_measure: ItemMeasure | None = None,
) -> None:
from headroom.proxy.outcome import RequestOutcome
@@ -1040,6 +1043,7 @@ class StreamingMixin:
waste_signals=waste_signals,
conversation_key=conversation_key,
conversation_tokens_saved=conversation_tokens_saved,
savings_measure=savings_measure,
)
await self._record_request_outcome(outcome)
@@ -1071,6 +1075,7 @@ class StreamingMixin:
session_key: str | None = None,
conversation_key: str | None = None,
conversation_tokens_saved: int | None = None,
savings_measure: ItemMeasure | None = None,
server_memory_tool_names: frozenset[str] | None = None,
client_beta: str | None = None,
) -> Response | StreamingResponse:
@@ -1118,6 +1123,7 @@ class StreamingMixin:
session_key=session_key,
conversation_key=conversation_key,
conversation_tokens_saved=conversation_tokens_saved,
savings_measure=savings_measure,
server_memory_tool_names=server_memory_tool_names,
client_beta=client_beta,
)
@@ -1152,6 +1158,7 @@ class StreamingMixin:
session_key: str,
conversation_key: str | None = None,
conversation_tokens_saved: int | None = None,
savings_measure: ItemMeasure | None = None,
server_memory_tool_names: frozenset[str] | None = None,
client_beta: str | None = None,
) -> Response | StreamingResponse:
@@ -1586,6 +1593,7 @@ class StreamingMixin:
status_code=upstream_response.status_code,
conversation_key=conversation_key,
conversation_tokens_saved=conversation_tokens_saved,
savings_measure=savings_measure,
)
self._cleanup_mid_turn_stream(session_key)
return Response(
@@ -1981,6 +1989,7 @@ class StreamingMixin:
),
conversation_key=conversation_key,
conversation_tokens_saved=conversation_tokens_saved,
savings_measure=savings_measure,
)
if supports_mid_turn_coalescing(client) and pending_messages:
pending_event = json.dumps(
+12
View File
@@ -33,6 +33,7 @@ from typing import Any
from headroom.log_safety import describe_exception
from headroom.proxy.conversation_savings import get_conversation_savings
from headroom.proxy.savings import ItemMeasure, get_item_ledger
from headroom.proxy.tool_schema_savings_policy import (
headline_tokens_saved,
tool_schema_saved_from_tags,
@@ -204,6 +205,11 @@ class RequestOutcome:
# path, where the outcome carries a per-turn delta.
conversation_key: str | None = None
conversation_tokens_saved: int | None = None
# savings_measure: the request's per-item original/sent counts against the
# conversation's previous request (``headroom.proxy.savings``), from
# which the per-request savings record is priced. None on paths that
# do not measure items yet.
savings_measure: ItemMeasure | None = None
transforms_applied: tuple[str, ...] = ()
waste_signals: dict[str, int] | None = None
num_messages: int = 0
@@ -339,6 +345,7 @@ class RequestOutcome:
conversation_key: str | None = None,
conversation_tokens_saved: int | None = None,
provider_input_tokens: int = 0,
savings_measure: ItemMeasure | None = None,
) -> RequestOutcome:
"""Construct an outcome from the locals available at streaming
finalize. Three streaming finalizers
@@ -414,6 +421,7 @@ class RequestOutcome:
tokens_saved=tokens_saved,
conversation_key=conversation_key,
conversation_tokens_saved=conversation_tokens_saved,
savings_measure=savings_measure,
attempted_input_tokens=optimized_tokens + tokens_saved,
status_code=status_code,
cache_read_tokens=cache_read_tokens,
@@ -539,6 +547,10 @@ async def emit_request_outcome(handler: Any, outcome: RequestOutcome) -> None:
# beacon must never add latency to, or take down, the request path.
record_outcome(outcome)
if outcome.savings_measure is not None:
# Only a request the provider served becomes the next one's baseline.
get_item_ledger().settle(outcome.savings_measure, succeeded=outcome.status_code < 400)
# A rejected turn must not feed the savings/cost/log success stats; that
# would let a request the provider never billed inflate the save-rate.
# Record it under the counter that names what happened, and stop.
+707
View File
@@ -0,0 +1,707 @@
"""One per-request savings record: counterfactual cost minus actual cost.
Every savings surface (/stats, the durable ledger, persistent savings, the
dashboard, the PERF line) is meant to derive from the :class:`SavingsRecord`
this module computes, so no surface prices savings its own way.
Two rulers meet here and are only ever combined through RATIOS:
* Headroom's tokenizer gives per-item counts of the request as the client sent
it (ORIGINAL) and as Headroom forwarded it (SENT). Both are counted with the
same tokenizer, so a tokenizer bias cancels in their difference.
* The provider's usage block gives exact billed totals (uncached, cache read,
cache write) but no per-item detail.
``k = billed / sent`` converts Headroom units into provider units, calibrated
on this request's own content. The counterfactual cache behaviour mirrors the
actual request's: removed tokens in items the previous request already carried
would have been cache reads to the extent the actual request hit its cache
(``warm``); removed tokens in new items would have been misses, written to
cache in the proportion the actual request wrote (``w``). Old items whose sent
copy Headroom changed since the previous request (``bust``) are misses that
Headroom caused, charged back against the saving.
Output tokens are identical in the counterfactual by construction; output
shaping is measured separately against its holdout and stays out of this
number.
``headroom.pricing`` is imported inside functions, as elsewhere on the proxy
path: its package ``__init__`` pulls in litellm (~4s), which must not land on
proxy startup.
"""
from __future__ import annotations
import hashlib
import json
import logging
import threading
from collections import OrderedDict
from collections.abc import Callable, Sequence
from dataclasses import dataclass, field, replace
from typing import TYPE_CHECKING, Any
from headroom.log_safety import describe_exception, safe_id
if TYPE_CHECKING:
from headroom.pricing.counterfactual import CacheMix, CacheRates
logger = logging.getLogger(__name__)
#: OpenAI caches prompts of at least 1024 tokens, in 128-token blocks.
OPENAI_MIN_CACHEABLE_TOKENS = 1024
OPENAI_CACHE_BLOCK_TOKENS = 128
#: Used when the catalog has no ``prompt_cache_min_tokens`` for a model.
DEFAULT_MIN_CACHEABLE_TOKENS = 1024
@dataclass(frozen=True, slots=True)
class ItemCounts:
"""One request item (tools block, system, or one input/message item).
Hashes identify the item's content; counts are in Headroom tokens. Lists of
items are aligned by position between ORIGINAL and SENT, so an item
Headroom dropped entirely keeps its slot with ``sent_hash=""`` and
``sent_tokens=0``. Tool-schema deferral is removal inside the tools item.
"""
original_hash: str
original_tokens: int
sent_hash: str
sent_tokens: int
@dataclass(frozen=True)
class ItemMeasure:
"""The Headroom-unit quantities the savings formula needs."""
#: ``S``: tokens of the whole sent body.
sent_tokens: int
#: ``R_old``: removed tokens (signed: an item Headroom grew counts against
#: it) in items within the longest prefix of the ORIGINAL body identical to
#: the previous ORIGINAL body.
removed_old: int
#: ``R_new``: removed tokens in every other item, signed likewise.
removed_new: int
#: ``P_S``: tokens of the longest prefix of the SENT body identical to the
#: previous SENT body.
sent_shared_prefix: int
#: Sent tokens of old items that fell outside the sent shared prefix: cache
#: misses caused only by Headroom changing what it had sent before.
bust: int
#: False when the old/new split is approximate (the totals stay exact):
#: Headroom changed the item count and the original/sent pairing had to be
#: approximated, or the baseline was lost (``baseline_lost``).
aligned: bool = True
#: True when the ledger lost what this conversation last sent (its
#: baseline or an in-flight request was evicted under the budget, or its
#: baseline may be truncated), so ``removed_new`` may hold old items. Such
#: a measure is also ``aligned=False``. Its record prices the new removal
#: at the cache-read rate, the lowest any input token costs, so the lost
#: knowledge can only understate the saving. A conversation's first
#: request is not lost: nothing was sent before, so everything is new.
baseline_lost: bool = False
#: Tokens Headroom added to new items (the negative part of
#: ``removed_new``), so the read-rate floor covers only what it removed.
grown_new: int = 0
#: Which measured request this is, for :meth:`ItemLedger.settle`.
ticket: int = field(default=0, compare=False)
@property
def removed(self) -> int:
return self.removed_old + self.removed_new
@dataclass(frozen=True)
class SavingsRecord:
"""What one request saved, in dollars and in both token rulers."""
saved_usd: float
actual_input_cost_usd: float
counterfactual_input_cost_usd: float
#: Removed tokens in Headroom units (``R``).
removed_tokens: int
#: The same removal in provider units (``k * R``).
saved_tokens: float
k: float
warm: float
write_share: float
#: True when usage was missing (aborted stream) or summed over server-side
#: iterations, so the cache split could not be read from this request.
estimated: bool = False
#: False when the catalog has no price for the model.
priced: bool = True
def _common_prefix_len(a: Sequence[str], b: Sequence[str]) -> int:
n = 0
for x, y in zip(a, b):
if x != y:
break
n += 1
return n
def measure_items(
items: Sequence[ItemCounts],
previous: Sequence[ItemCounts] | None,
) -> ItemMeasure:
"""Split one request's removal into old/new items against the previous request.
``previous`` is the same conversation's previous request, or ``None`` on its
first request, where everything is new and nothing is shared.
"""
n_old = _common_prefix_len(
[i.original_hash for i in items], [i.original_hash for i in previous or ()]
)
n_sent_shared = _common_prefix_len(
[i.sent_hash for i in items], [i.sent_hash for i in previous or ()]
)
removed = [i.original_tokens - i.sent_tokens for i in items]
sent = [max(0, i.sent_tokens) for i in items]
return ItemMeasure(
sent_tokens=sum(sent),
removed_old=sum(removed[:n_old]),
removed_new=sum(removed[n_old:]),
sent_shared_prefix=sum(sent[:n_sent_shared]),
bust=sum(sent[n_sent_shared:n_old]),
grown_new=-sum(r for r in removed[n_old:] if r < 0),
)
def _dump(value: Any) -> str:
return json.dumps(value, ensure_ascii=False, separators=(",", ":"), default=str)
#: One request item as (identity text, billable text): the identity is hashed
#: to tell items apart, the billable text is what gets tokenized.
Item = tuple[str, str]
def _item(value: Any) -> Item:
identity = _dump(value)
if (
isinstance(value, dict)
and value.get("type") == "reasoning"
and "encrypted_content" in value
):
# The provider bills the reasoning it decrypts, not the ciphertext,
# which would otherwise be most of a long Codex transcript's count.
billable = _dump({k: v for k, v in value.items() if k != "encrypted_content"})
return identity, billable
return identity, identity
def responses_items(body: dict[str, Any]) -> list[Item]:
"""A ``/v1/responses`` body as items in prompt order.
Tools, then instructions, then each ``input`` item: the order the provider
caches the prompt in, which is what the shared-prefix comparison needs.
Codex's ``additional_tools`` input items stay where the client put them.
"""
items = [_item(body.get("tools") or []), _item(body.get("instructions") or "")]
# OpenAI-compatible upstreams accept ``messages`` in place of ``input``.
raw = body.get("input") if body.get("input") is not None else body.get("messages")
if isinstance(raw, list):
items.extend(_item(item) for item in raw)
elif raw:
items.append(_item(raw))
return items
def _digest(text: str) -> str:
return hashlib.blake2b(text.encode("utf-8", errors="replace"), digest_size=8).hexdigest()
#: Items kept across all conversations' baselines and requests awaiting their
#: outcome, combined. A long Codex session carries 1000+ items; bounding items,
#: not conversations, bounds memory: up to twice this while one conversation's
#: baseline and its in-flight request exceed it (see :class:`ItemLedger`).
MAX_STORED_ITEMS = 100_000
#: Conversations remembered as having lost their baseline, a memory bound
#: (one short key each). Past it the ledger can no longer tell a first request
#: from a lost one and treats every request without a baseline as lost: priced
#: at the cache-read floor, which can understate a saving but never overstate it.
MAX_LOST_KEYS = 100_000
def _align(
original: list[tuple[str, int]],
sent: list[tuple[str, int]],
previous: Sequence[ItemCounts] | None,
) -> tuple[list[ItemCounts], bool]:
"""Pair original and sent items, and say whether the pairing is exact.
Same item count: by position (compression rewrites items in place). A
different count (a hook folded or dropped items, or Headroom inserted
some): identical items at either end pair up; the unmatched originals in
between are removed, and the unmatched sent items in between are folded
into the first of their slots. The totals stay exact either way; only
which items count as old is approximate, hence ``aligned=False``. An item
compressed in place still pairs at the head when the previous request
paired the same original with the same sent copy at that position.
Equal counts pair by position even if a fold and an insert cancel out.
"""
if len(original) == len(sent):
return [ItemCounts(oh, ot, sh, st) for (oh, ot), (sh, st) in zip(original, sent)], True
limit = min(len(original), len(sent))
prev = previous or ()
def same_as_before(i: int) -> bool:
return (
i < len(prev)
and original[i][0] == prev[i].original_hash
and sent[i][0] == prev[i].sent_hash
)
head = 0
while head < limit and (original[head][0] == sent[head][0] or same_as_before(head)):
head += 1
tail = 0
while (
tail < limit - head
and original[len(original) - 1 - tail][0] == sent[len(sent) - 1 - tail][0]
):
tail += 1
mid_original = original[head : len(original) - tail]
mid_sent = sent[head : len(sent) - tail]
if len(mid_sent) == 1:
folded_hash = mid_sent[0][0]
else:
folded_hash = _digest("\x00".join(h for h, _ in mid_sent)) if mid_sent else ""
folded_tokens = sum(t for _, t in mid_sent)
middle = [ItemCounts(oh, ot, "", 0) for oh, ot in mid_original] or [ItemCounts("", 0, "", 0)]
middle[0] = ItemCounts(
middle[0].original_hash, middle[0].original_tokens, folded_hash, folded_tokens
)
pairs = [ItemCounts(oh, ot, sh, st) for (oh, ot), (sh, st) in zip(original[:head], sent[:head])]
pairs += middle
pairs += [ItemCounts(h, t, h, t) for h, t in original[len(original) - tail :]]
return pairs, False
class ItemLedger:
"""The previous request of each conversation, for the old/new item split.
One instance per process. A request is measured against the conversation's
last SUCCESSFUL request and becomes the baseline only once its own outcome
succeeds (:meth:`settle`): a failed call never reached the provider's cache.
Each measure carries a ticket, issued in measure order, so a slow earlier
turn settling late cannot replace a newer turn's baseline. Items re-sent
from the baseline reuse its token counts, so a long transcript is
tokenized once rather than every turn.
Baselines and requests awaiting their outcome share one budget,
:data:`MAX_STORED_ITEMS`. Past it, whichever is least recent goes first,
by ticket: a pending request by its own ticket (its slot is reserved when
the ticket is issued, so a slow measure keeps its place), a baseline by
the last ticket that used it. Settling counts as a use. A request being
admitted never evicts its own conversation's baseline: that baseline is
replaced only when a newer request for the conversation succeeds. When
only that baseline and one pending request are left, both are kept, even
past the budget, so a conversation larger than half the budget still
advances; memory then peaks at twice :data:`MAX_STORED_ITEMS` items.
A request is kept only up to its first :data:`MAX_STORED_ITEMS` items: the
comparison is a prefix, so the next request still gets the shared-prefix
credit for them.
Whatever is evicted, its conversation is remembered as lost, so its next
request is marked ``baseline_lost`` (and ``aligned=False``): its record
prices the new removal at the cache-read floor instead of as new, at the
miss rate. So is a request measured against a baseline of exactly
:data:`MAX_STORED_ITEMS` items, which may be truncated. A conversation's
first request is new, not lost. The memory of losses is bounded by
:data:`MAX_LOST_KEYS`; past it, every request without a baseline is
treated as lost, and a forgotten loss's stale baseline is dropped with it.
"""
def __init__(self) -> None:
self._lock = threading.Lock()
# key -> (settled ticket, items, last ticket that used it), oldest use first.
self._previous: OrderedDict[str, tuple[int, list[ItemCounts], int]] = OrderedDict()
# ticket -> (key, items), in ticket order.
self._pending: OrderedDict[int, tuple[str, list[ItemCounts]]] = OrderedDict()
# key -> newest ticket whose items were evicted, oldest loss first.
self._lost: OrderedDict[str, int] = OrderedDict()
# True once a loss was forgotten: a missing baseline may then be a lost one.
self._lost_overflowed = False
self._stored = 0
self._pending_items = 0
self._seq = 0
def measure(
self,
key: str,
original: Sequence[Item],
sent: Sequence[Item],
count: Callable[[str], int],
) -> ItemMeasure:
"""Measure one request against ``key``'s baseline; settle it later."""
with self._lock:
self._seq += 1
seq = self._seq
entry = self._previous.get(key)
if entry is not None:
self._previous[key] = (entry[0], entry[1], seq)
self._previous.move_to_end(key)
lost = key in self._lost or (entry is None and self._lost_overflowed)
self._pending[seq] = (key, [])
previous = entry[1] if entry else None
known: dict[str, int] = {}
for item in previous or ():
known[item.original_hash] = item.original_tokens
known[item.sent_hash] = item.sent_tokens
def counted(items: Sequence[Item]) -> list[tuple[str, int]]:
out = []
for identity, billable in items:
digest = _digest(identity)
tokens = known.get(digest)
if tokens is None:
tokens = known[digest] = int(count(billable))
out.append((digest, tokens))
return out
try:
items, aligned = _align(counted(original), counted(sent), previous)
except BaseException:
# The caller logs and never settles: release the reserved slot.
with self._lock:
self._pending.pop(seq, None)
raise
if previous is not None and len(previous) == MAX_STORED_ITEMS:
# Possibly a truncated baseline: items past it may be old.
lost = True
with self._lock:
if seq in self._pending: # not already evicted while counting
kept = items[:MAX_STORED_ITEMS]
self._pending[seq] = (key, kept)
self._pending_items += len(kept)
self._evict(key)
return replace(
measure_items(items, previous),
aligned=aligned and not lost,
baseline_lost=lost,
ticket=seq,
)
def settle(self, measure: ItemMeasure, *, succeeded: bool) -> None:
"""Make a measured request its conversation's baseline if it succeeded."""
seq = measure.ticket
with self._lock:
pending = self._pending.pop(seq, None)
if pending is None:
return
key, items = pending
self._pending_items -= len(items)
if not succeeded:
return
current = self._previous.get(key)
if current is not None and current[0] > seq:
return
# Moving items from pending to stored never grows the total.
self._stored -= len(current[1]) if current else 0
self._previous[key] = (seq, items, self._seq)
self._previous.move_to_end(key)
self._stored += len(items)
if self._lost.get(key, seq) < seq:
# Newer than everything evicted: the baseline is whole again.
del self._lost[key]
def _evict(self, admitting: str) -> None:
"""Forget the least recent entries until the combined budget holds.
Called with the lock held, right after admitting a request for
``admitting``, whose baseline is never a candidate. Each pass removes
one entry until only that baseline and one pending request are left:
each holds at most :data:`MAX_STORED_ITEMS` items, so they stay.
"""
while self._stored + self._pending_items > MAX_STORED_ITEMS:
oldest_ticket = next(iter(self._pending), None)
oldest_key = next((k for k in self._previous if k != admitting), None)
if oldest_key is None and len(self._pending) <= 1:
return
# On a tie the baseline goes: the request that used it is newer.
if oldest_key is not None and (
oldest_ticket is None or self._previous[oldest_key][2] <= oldest_ticket
):
settled, evicted, _ = self._previous.pop(oldest_key)
self._stored -= len(evicted)
self._mark_lost(oldest_key, settled)
else:
ticket, (key, dropped) = self._pending.popitem(last=False)
self._pending_items -= len(dropped)
self._mark_lost(key, ticket)
def _mark_lost(self, key: str, ticket: int) -> None:
"""Remember that ``key``'s items up to ``ticket`` were evicted (lock held)."""
self._lost[key] = max(ticket, self._lost.get(key, 0))
self._lost.move_to_end(key)
while len(self._lost) > MAX_LOST_KEYS:
forgotten, _ = self._lost.popitem(last=False)
if not self._lost_overflowed:
self._lost_overflowed = True
logger.info(
"savings: lost-baseline memory full (MAX_LOST_KEYS=%d); requests "
"without a baseline now price at the cache-read floor",
MAX_LOST_KEYS,
)
# Unmarked, a stale baseline would price old items as new: drop it,
# so the conversation counts as lost under the overflow rule.
stale = self._previous.pop(forgotten, None)
if stale is not None:
self._stored -= len(stale[1])
_ledger = ItemLedger()
def get_item_ledger() -> ItemLedger:
"""Process-wide ledger: one conversation spans many requests and handlers."""
return _ledger
def prefix_key(conversation: str | None, model: str, original: Sequence[Item]) -> str:
"""Which previous request a request is compared against.
The conversation's explicit identity (``savings_conversation_key``) plus
the tools and instructions: a side request in the same session (compaction,
review) carries a different prefix and must not replace the conversation's
previous request. Without an identity, the model plus the first input item
stand in. Requests that collide on the key overwrite each other's previous
request, shortening the shared prefix: more of the removal is then priced
as new, at the miss rate, which overstates the saving.
"""
head = "\x00".join(identity for identity, _ in original[:2])
if conversation:
# The model too: the provider's cache is per model, and so are counts.
return _digest("\x00".join([conversation, model, head]))
return _digest("\x00".join([model, *(identity for identity, _ in original[:3])]))
def _input_cost(
rates: CacheRates, *, uncached: float, read: float, write_5m: float, write_1h: float
) -> float:
return (
uncached * rates.uncached
+ read * rates.read
+ write_5m * rates.write_5m
+ write_1h * rates.write_1h
)
def _min_cacheable_tokens(model: str, provider: str | None) -> int:
if provider == "openai":
return OPENAI_MIN_CACHEABLE_TOKENS
from headroom.pricing.counterfactual import catalog_min_cacheable_tokens
return catalog_min_cacheable_tokens(model) or DEFAULT_MIN_CACHEABLE_TOKENS
def _unpriced(measure: ItemMeasure, *, k: float, estimated: bool) -> SavingsRecord:
return SavingsRecord(
saved_usd=0.0,
actual_input_cost_usd=0.0,
counterfactual_input_cost_usd=0.0,
removed_tokens=measure.removed,
saved_tokens=k * measure.removed,
k=k,
warm=0.0,
write_share=0.0,
estimated=estimated,
priced=False,
)
def compute_savings_record(
measure: ItemMeasure,
usage: CacheMix | None,
*,
model: str,
provider: str | None = None,
estimated: bool = False,
) -> SavingsRecord:
"""Price one request's removal as ``counterfactual_cost - actual_cost``.
``usage`` is the provider's billed input as a :class:`CacheMix` (normalized
here, so an inferred OpenAI write counts as the uncached slice it is).
``None`` or an empty mix means the provider reported nothing (an aborted
stream); the record then prices at list with ``k = 1`` and is marked
``estimated``.
For a request the provider ran as several server-side iterations, pass the
FIRST iteration's usage when it is available (``estimated=False``); the
later iterations are a cost the caller books separately. Pass the summed
usage with ``estimated=True`` only when per-iteration usage is not
available: the sum describes more prompts than the one sent body, so
``k`` is not calibrated from it (``k = 1``) and only its cache ratios are
used. ``provider`` must be given for OpenAI's 128-token cache blocks to
apply.
Never raises: a savings figure must not be able to fail a request.
"""
try:
record = _compute(measure, usage, model=model, provider=provider, estimated=estimated)
# An approximate item pairing leaves the totals exact but not the
# old/new split the price depends on.
return record if measure.aligned else replace(record, estimated=True)
except Exception as exc: # noqa: BLE001 - pricing must never fail a request
logger.debug(
"savings: record failed for model=%s: %s", safe_id(model), describe_exception(exc)
)
return _unpriced(measure, k=1.0, estimated=True)
def _compute(
measure: ItemMeasure,
usage: CacheMix | None,
*,
model: str,
provider: str | None,
estimated: bool,
) -> SavingsRecord:
from headroom.pricing.counterfactual import LONG_CONTEXT_THRESHOLD_TOKENS, resolve_rates
def rates_for(billed: float, *, for_billing: bool = False) -> CacheRates | None:
return resolve_rates(
model,
long_context=billed > LONG_CONTEXT_THRESHOLD_TOKENS,
provider=provider,
for_billing=for_billing,
)
removed = measure.removed
# With the baseline lost, the new items may be old ones the provider had
# cached. Their removal prices at the read rate, the floor of any input
# token: the cache ratio and minimum below describe a prefix the ledger no
# longer knows. Tokens Headroom added to new items stay at the miss rate.
floor = measure.removed_new + measure.grown_new if measure.baseline_lost else 0
mix = usage.normalized() if usage is not None else None
if mix is None or mix.billed <= 0:
actual_rates = rates_for(measure.sent_tokens)
cf_rates = rates_for(measure.sent_tokens + removed)
if actual_rates is None or cf_rates is None:
return _unpriced(measure, k=1.0, estimated=True)
actual = measure.sent_tokens * actual_rates.uncached
counterfactual = (
measure.sent_tokens + removed - floor
) * cf_rates.uncached + floor * cf_rates.read
return SavingsRecord(
saved_usd=counterfactual - actual,
actual_input_cost_usd=actual,
counterfactual_input_cost_usd=counterfactual,
removed_tokens=removed,
saved_tokens=float(removed),
k=1.0,
warm=0.0,
write_share=0.0,
estimated=True,
)
billed = mix.billed
writes = mix.write_5m + mix.write_1h
k = billed / measure.sent_tokens if measure.sent_tokens > 0 and not estimated else 1.0
hit = mix.read / billed
sigma = measure.sent_shared_prefix / measure.sent_tokens if measure.sent_tokens > 0 else 0.0
warm = min(1.0, hit / sigma) if sigma > 0 else 0.0
misses = mix.uncached + writes
w = writes / misses if misses > 0 else 0.0
# The counterfactual must obey its own cache minimum. The catalog's
# minimum is used unless this request was demonstrably cached or written
# below it; then the catalog is wrong and no minimum is applied, the
# direction that cannot inflate the saving.
min_cacheable = _min_cacheable_tokens(model, provider)
proven = mix.read if mix.read > 0 else (billed if writes > 0 else min_cacheable)
threshold = min_cacheable if proven >= min_cacheable else 0
# Removed or busted old tokens could only have been reads if the
# counterfactual's old prefix was itself cacheable.
raw_warm = warm
if k * (measure.sent_shared_prefix + measure.removed_old) < threshold:
warm = 0.0
# The counterfactual's buckets are built whole, not as deltas clamped onto
# the actual ones: a counterfactual smaller than the request (an old item
# Headroom grew) must not inherit reads or writes it was too small to make.
# Its reads are the actual reads plus the removed/busted old tokens that
# would have hit, minus the reads a grown old item contributed (those go
# whatever the prefix size), never more than its whole billed input, and
# none below its cache minimum. ``cf_billed`` is ``k * original`` when k is
# calibrated; the clamp matters only for summed (estimated) usage.
cf_billed = max(0.0, billed + k * removed)
gained = k * (max(0, measure.removed_old) + measure.bust) * warm
lost = k * min(0, measure.removed_old) * raw_warm
cf_read = min(max(0.0, mix.read + gained + lost), cf_billed)
if cf_read < threshold:
# Never fires when nothing changed: the actual reads are 0 or proven
# at least the threshold.
cf_read = 0.0
if provider == "openai" and cf_read != mix.read:
# Hits come in whole 128-token blocks: round down, but a counterfactual
# that adds reads keeps at least the actual ones. The ``!=`` keeps an
# unchanged request exactly at the actual reads.
blocks = (cf_read // OPENAI_CACHE_BLOCK_TOKENS) * OPENAI_CACHE_BLOCK_TOKENS
cf_read = max(blocks, mix.read) if cf_read > mix.read else blocks
if floor > 0:
cf_read = min(cf_read + k * floor, cf_billed)
# Everything else missed, written in the proportion the request wrote,
# unless the counterfactual is smaller than anything shown cacheable: a
# write it could not make would inflate the saving.
cf_miss = cf_billed - cf_read
cf_write = cf_miss * w if cf_billed >= min(min_cacheable, proven) else 0.0
w5_share = mix.write_5m / writes if writes > 0 else 1.0
cf_buckets = {
"uncached": cf_miss - cf_write,
"read": cf_read,
"write_5m": cf_write * w5_share,
"write_1h": cf_write * (1.0 - w5_share),
}
actual_buckets = {
"uncached": float(mix.uncached),
"read": float(mix.read),
"write_5m": float(mix.write_5m),
"write_1h": float(mix.write_1h),
}
# Tokens the request was billed for price at billing rates on both sides
# (the counterfactual's at its own context tier); only the tokens a
# counterfactual bucket holds beyond the actual one price at the
# savings-estimate rates. Every priced quantity is nonnegative.
actual_rates = rates_for(billed, for_billing=True)
cf_billing_rates = rates_for(cf_billed, for_billing=True)
cf_rates = rates_for(cf_billed)
if actual_rates is None or cf_billing_rates is None or cf_rates is None:
return _unpriced(measure, k=k, estimated=estimated)
actual = _input_cost(actual_rates, **actual_buckets)
counterfactual = _input_cost(
cf_billing_rates, **{b: min(n, actual_buckets[b]) for b, n in cf_buckets.items()}
) + _input_cost(cf_rates, **{b: max(0.0, n - actual_buckets[b]) for b, n in cf_buckets.items()})
return SavingsRecord(
saved_usd=counterfactual - actual,
actual_input_cost_usd=actual,
counterfactual_input_cost_usd=counterfactual,
removed_tokens=removed,
saved_tokens=k * removed,
k=k,
warm=warm,
write_share=w,
estimated=estimated,
)
__all__ = [
"Item",
"ItemCounts",
"ItemLedger",
"ItemMeasure",
"SavingsRecord",
"compute_savings_record",
"get_item_ledger",
"measure_items",
"prefix_key",
"responses_items",
]
+376
View File
@@ -0,0 +1,376 @@
"""Per-item original/sent measurement feeding the savings record."""
from __future__ import annotations
import json
import threading
from typing import Any
import httpx
import pytest
import respx
from fastapi.testclient import TestClient
from headroom.proxy import savings
from headroom.proxy.loopback_guard import require_loopback
from headroom.proxy.savings import ItemLedger, ItemMeasure, prefix_key, responses_items
from headroom.proxy.server import ProxyConfig, create_app
def _words(text: str) -> int:
return len(text.split(","))
def _items(*texts: str) -> list[tuple[str, str]]:
return [(t, t) for t in texts]
def _turn(ledger, original, sent, count=_words, *, succeeded=True):
"""Measure one request and settle it, as the outcome funnel does."""
m = ledger.measure("conv", original, sent, count)
ledger.settle(m, succeeded=succeeded)
return m
def test_a_conversation_is_measured_against_its_previous_request():
ledger = ItemLedger()
calls: list[str] = []
def count(text: str) -> int:
calls.append(text)
return _words(text)
first = ['["t1","t2"]', '"sys"', '{"a":1,"b":2,"c":3}']
sent1 = ['["t1"]', '"sys"', '{"a":1,"b":2,"c":3}']
second = [*first, '{"d":4,"e":5}']
sent2 = [*sent1, '{"d":4}']
assert _turn(ledger, _items(*first), _items(*sent1), count) == ItemMeasure(
sent_tokens=5, removed_old=0, removed_new=1, sent_shared_prefix=0, bust=0
)
assert _turn(ledger, _items(*second), _items(*sent2), count) == ItemMeasure(
sent_tokens=6, removed_old=1, removed_new=1, sent_shared_prefix=5, bust=0
)
# Re-sent items are tokenized once, not once per turn.
assert len(calls) == len(set(calls))
def test_items_follow_the_providers_prompt_order():
body = {"tools": [{"name": "t"}], "instructions": "be brief", "input": [{"x": 1}, {"y": 2}]}
assert responses_items(body) == _items('[{"name":"t"}]', '"be brief"', '{"x":1}', '{"y":2}')
assert responses_items({"input": "hi"}) == _items("[]", '""', '"hi"')
def test_reasoning_ciphertext_is_identity_not_billed_text():
"""Codex replays encrypted reasoning every turn; counting the ciphertext
made Headroom's count drift far above the provider's bill."""
item = {"type": "reasoning", "summary": [], "encrypted_content": "gAAAA" * 50}
[(identity, billable)] = responses_items({"input": [item]})[2:]
assert "gAAAA" in identity
assert billable == '{"type":"reasoning","summary":[]}'
def test_a_folded_transcript_keeps_the_old_items_old():
"""A hook folding two items into one shifts every later position; pairing
by position then priced untouched new items as removed."""
ledger = ItemLedger()
t, i = "[]", '"sys"'
a, b, c, d = ("," * 9 + x for x in "abcd") # 10 tokens each under _words
_turn(ledger, _items(t, i, a, b, c), _items(t, i, a, b, c), _words)
m = _turn(ledger, _items(t, i, a, b, c, d), _items(t, i, "f", c, d), _words)
assert (m.removed_old, m.removed_new) == (19, 0)
# The fold moved c: its sent copy is outside the shared prefix, a bust.
assert m.bust == 11 and m.aligned is False
def test_a_fold_after_an_item_compressed_in_place_keeps_the_prefix():
"""The common Codex case: an early tool output compressed every turn, then
a hook folds the new items. The compressed item must still pair at the head,
or every later item is priced as removed and as a bust."""
ledger = ItemLedger()
t, i = "[]", '"sys"'
big, small = "," * 99 + "x", "x" # 100 and 1 tokens under _words
n1, n2 = "," * 9 + "n", "," * 9 + "m"
_turn(ledger, _items(t, i, big), _items(t, i, small), _words)
m = _turn(ledger, _items(t, i, big, n1, n2), _items(t, i, small, "f"), _words)
assert (m.sent_shared_prefix, m.bust) == (3, 0)
assert (m.removed_old, m.removed_new) == (99, 19)
assert m.aligned is False
def test_an_approximate_pairing_marks_the_record_estimated():
m = ItemMeasure(
sent_tokens=10, removed_old=0, removed_new=5, sent_shared_prefix=0, bust=0, aligned=False
)
assert savings.compute_savings_record(m, None, model="unpriced-model").estimated
def test_a_failed_request_does_not_become_the_baseline():
ledger = ItemLedger()
first = _items("[]", '"sys"', '"a"')
_turn(ledger, first, first)
_turn(ledger, _items("[]", '"other"', '"b"'), _items("[]", '"other"', '"b"'), succeeded=False)
assert _turn(ledger, first, first).sent_shared_prefix == 3
def test_a_late_settling_older_turn_keeps_the_newer_baseline():
ledger = ItemLedger()
old, new = _items("[]", '"a"'), _items("[]", '"a"', '"b"')
slow = ledger.measure("conv", old, old, _words)
fast = ledger.measure("conv", new, new, _words)
ledger.settle(fast, succeeded=True)
ledger.settle(slow, succeeded=True)
assert _turn(ledger, new, new).sent_shared_prefix == 3
def test_requests_awaiting_their_outcome_are_bounded_by_items(monkeypatch):
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
abandoned = ledger.measure("a", _items("[]", '"x"', '"y"'), _items("[]", '"x"', '"y"'), _words)
kept = ledger.measure("b", _items("[]", '"z"'), _items("[]", '"z"'), _words)
# The oldest pending request was forgotten; settling it is a no-op.
ledger.settle(abandoned, succeeded=True)
ledger.settle(kept, succeeded=True)
# Each probe is itself a pending request under the budget: kept first.
assert (
ledger.measure("b", _items("[]", '"z"'), _items("[]", '"z"'), _words).sent_shared_prefix
== 2
)
again = _items("[]", '"x"', '"y"')
assert ledger.measure("a", again, again, _words).sent_shared_prefix == 0
def test_a_slow_older_measure_never_evicts_a_newer_one(monkeypatch):
"""JerrettDavis's #4240 interleaving: ticket 1 stalls inside token
counting while ticket 2 measures four items. Eviction must drop the older
ticket, not the newer one that finished first."""
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
counting, release = threading.Event(), threading.Event()
def stalled(text: str) -> int:
counting.set()
assert release.wait(timeout=10)
return _words(text)
older: list[ItemMeasure] = []
slow = _items("[]", '"x"')
thread = threading.Thread(target=lambda: older.append(ledger.measure("a", slow, slow, stalled)))
thread.start()
assert counting.wait(timeout=10)
new = _items("[]", '"b1"', '"b2"', '"b3"')
newer = ledger.measure("b", new, new, _words)
release.set()
thread.join(timeout=10)
ledger.settle(newer, succeeded=True)
ledger.settle(older[0], succeeded=True)
assert ledger.measure("b", new, new, _words).sent_shared_prefix == 4
def test_settled_and_pending_items_share_one_budget(monkeypatch):
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
idle = _items("[]", '"a1"', '"a2"', '"a3"')
_turn(ledger, idle, idle)
active = _items("[]", '"b1"', '"b2"', '"b3"')
in_flight = ledger.measure("b", active, active, _words)
# Admitting the in-flight request evicted the idle baseline at once.
assert ledger._stored + ledger._pending_items == 4
ledger.settle(in_flight, succeeded=True)
assert ledger.measure("b", active, active, _words).sent_shared_prefix == 4
assert ledger.measure("conv", idle, idle, _words).sent_shared_prefix == 0
def test_a_request_larger_than_the_budget_is_kept_up_to_the_budget(monkeypatch):
"""Keeping nothing would leave the next request with no baseline. The
first items are kept instead; a request measured against them may have
old items past the cut, so its baseline counts as lost. Each turn still
replaces the baseline, however large the conversation."""
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
big = _items("[]", '"s"', '"u1"', '"u2"', '"u3"')
m = _turn(ledger, big, big)
assert (m.aligned, m.baseline_lost) == (True, False)
longer = [*big, *_items('"u4"')]
m = _turn(ledger, longer, longer)
assert (m.sent_shared_prefix, m.aligned, m.baseline_lost) == (4, False, True)
assert not ledger._lost # the cut-off rule fired, nothing was evicted
assert ledger._stored + ledger._pending_items <= 4
def test_a_failed_turn_keeps_its_conversations_baseline(monkeypatch):
"""Devin's #4240 example: admitting a four-item turn under a four-item
budget evicted the conversation's own baseline; the turn then got a 429,
and the next identical request saw no shared prefix at all."""
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
turn = _items("[]", '"s"', '"u1"', '"u2"')
_turn(ledger, turn, turn)
_turn(ledger, turn, turn, succeeded=False)
assert _turn(ledger, turn, turn).sent_shared_prefix == 4
def test_overlapping_turns_keep_the_baseline_until_one_succeeds(monkeypatch):
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 8)
ledger = ItemLedger()
base = _items("[]", '"s"', '"u1"')
_turn(ledger, base, base)
first = [*base, *_items('"u2"')]
second = [*first, *_items('"u3"', '"u4"')]
m1 = ledger.measure("conv", first, first, _words)
# Three baseline items plus ten pending: the older turn goes, not the baseline.
m2 = ledger.measure("conv", second, second, _words)
ledger.settle(m2, succeeded=False)
ledger.settle(m1, succeeded=False)
m3 = ledger.measure("conv", first, first, _words)
# Still the old baseline, but the forgotten turn may have been cached.
assert (m3.sent_shared_prefix, m3.baseline_lost, m3.aligned) == (3, True, False)
ledger.settle(m3, succeeded=True)
m4 = _turn(ledger, second, second)
assert (m4.sent_shared_prefix, m4.baseline_lost, m4.aligned) == (4, False, True)
def _probe(ledger, key, items):
m = ledger.measure(key, items, items, _words)
ledger.settle(m, succeeded=True)
return m
def test_a_lost_baseline_is_flagged_and_a_new_conversation_is_not(monkeypatch):
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
ledger = ItemLedger()
a, b, c = (_items("[]", f'"{n}1"', f'"{n}2"') for n in "abc")
assert not _probe(ledger, "a", a).baseline_lost
# Admitting b evicts a's baseline: another conversation, the least recent.
assert not _probe(ledger, "b", b).baseline_lost
again = _probe(ledger, "a", a)
assert (again.sent_shared_prefix, again.baseline_lost, again.aligned) == (0, True, False)
assert not _probe(ledger, "c", c).baseline_lost
def test_past_the_loss_bound_every_missing_baseline_counts_as_lost(monkeypatch):
"""Forgetting a loss must not let its conversation pass for a new one."""
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
monkeypatch.setattr(savings, "MAX_LOST_KEYS", 1)
ledger = ItemLedger()
a, b, c, d = (_items("[]", f'"{n}1"', f'"{n}2"') for n in "abcd")
_probe(ledger, "a", a)
_probe(ledger, "b", b) # loses a
_probe(ledger, "c", c) # loses b; a's loss is forgotten
assert _probe(ledger, "a", a).baseline_lost
assert _probe(ledger, "d", d).baseline_lost
def test_a_forgotten_loss_drops_the_stale_baseline_it_leaves(monkeypatch):
"""a's newer request was evicted while its baseline survived. Once that
loss is forgotten, the stale baseline must not pass for a whole one, or
the evicted request's cached items would price as new."""
monkeypatch.setattr(savings, "MAX_STORED_ITEMS", 4)
monkeypatch.setattr(savings, "MAX_LOST_KEYS", 1)
ledger = ItemLedger()
one, two = _items('"x"'), _items("[]", '"y"')
_probe(ledger, "x", one)
_probe(ledger, "a", one)
a1 = ledger.measure("a", one, one, _words)
ledger.settle(ledger.measure("x", one, one, _words), succeeded=False) # x used after a1
a2 = ledger.measure("a", one, one, _words)
ledger.measure("b", one, one, _words) # evicts a1: a is lost, its baseline stays
ledger.settle(a1, succeeded=True)
ledger.settle(a2, succeeded=False)
assert "a" in ledger._lost
ledger.measure("c", two, two, _words) # evicts x's baseline: a's loss is forgotten
m = ledger.measure("a", one, one, _words)
assert (m.baseline_lost, m.aligned) == (True, False)
def test_a_measure_that_fails_releases_its_slot():
"""The caller logs a failed measure and never settles it."""
ledger = ItemLedger()
def broken(text: str) -> int:
raise ValueError("tokenizer failed")
for _ in range(3):
with pytest.raises(ValueError):
ledger.measure("conv", _items("[]"), _items("[]"), broken)
assert not ledger._pending
def test_messages_stand_in_for_input():
assert responses_items({"messages": [{"role": "user", "content": "hi"}]})[2:] == _items(
'{"role":"user","content":"hi"}'
)
def test_side_requests_do_not_replace_the_conversations_previous_request():
main = _items("[]", '"agent"', '"hi"')
side = _items("[]", '"summarize"', '"hi"')
assert prefix_key("conv-1", "m", main) != prefix_key("conv-1", "m", side)
assert prefix_key("conv-1", "m", main) == prefix_key(
"conv-1", "m", [*main, ('"more"', '"more"')]
)
assert prefix_key(None, "m", main) == prefix_key(None, "m", [*main, ('"more"', '"more"')])
assert prefix_key(None, "m", main) != prefix_key(None, "other", main)
def _sse(usage: dict) -> str:
done = {"type": "response.completed", "response": {"id": "r", "output": [], "usage": usage}}
return f"event: response.completed\ndata: {json.dumps(done)}\n\n"
@pytest.mark.parametrize("stream", [False, True])
@respx.mock
def test_responses_outcomes_carry_the_measure(monkeypatch, stream):
app = create_app(
ProxyConfig(
optimize=False,
cache_enabled=False,
rate_limit_enabled=False,
cost_tracking_enabled=False,
log_requests=False,
)
)
app.dependency_overrides[require_loopback] = lambda: None
outcomes: list[Any] = []
# Capture at the funnel's first effect, so the real funnel (which settles
# the measure) still runs.
monkeypatch.setattr("headroom.telemetry.session.record_outcome", outcomes.append)
monkeypatch.setattr(savings, "_ledger", ItemLedger())
usage = {"input_tokens": 10, "output_tokens": 1}
respx.post("https://api.openai.com/v1/responses").mock(
return_value=httpx.Response(
200,
text=_sse(usage),
headers={"content-type": "text/event-stream"},
)
if stream
else httpx.Response(200, json={"id": "r", "output": [], "usage": usage})
)
turn1 = [{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "a"}]}]
turn2 = [
*turn1,
{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "b"}]},
]
with TestClient(app) as client:
for items in (turn1, turn2):
r = client.post(
"/v1/responses",
json={
"model": "gpt-4o",
"instructions": "be brief",
"input": items,
"stream": stream,
"client_metadata": {"session_id": "conv-test"},
},
headers={"authorization": "Bearer sk-test"},
)
assert r.status_code == 200
r.read()
first, second = (o.savings_measure for o in outcomes)
assert first.removed == 0 and first.sent_shared_prefix == 0
# Nothing was compressed, so the whole first request is the shared prefix.
assert second.sent_shared_prefix == first.sent_tokens
assert second.removed == 0 and second.bust == 0
+368
View File
@@ -0,0 +1,368 @@
"""The per-request savings record: counterfactual cost minus actual cost."""
from __future__ import annotations
import dataclasses
import pytest
from headroom.pricing import counterfactual
from headroom.pricing.counterfactual import CacheMix, CacheRates
from headroom.proxy.savings import (
ItemCounts,
ItemMeasure,
compute_savings_record,
measure_items,
)
# $3/M input, $0.30/M cache read, $3.75/M 5m write, $6/M 1h write; the
# long-context tier doubles everything.
IN, READ, WRITE, WRITE_1H = 3e-6, 3e-7, 3.75e-6, 6e-6
BASE = CacheRates(read=READ, write_5m=WRITE, write_1h=WRITE_1H, uncached=IN)
LONG = CacheRates(read=2 * READ, write_5m=2 * WRITE, write_1h=2 * WRITE_1H, uncached=2 * IN)
@pytest.fixture(autouse=True)
def _fixed_catalog(monkeypatch):
"""Pin the catalog so the arithmetic is tested whatever litellm ships."""
def rates(model, *, long_context=False, provider=None, for_billing=False):
return None if model == "unpriced" else (LONG if long_context else BASE)
monkeypatch.setattr(counterfactual, "resolve_rates", rates)
monkeypatch.setattr(
counterfactual,
"catalog_min_cacheable_tokens",
lambda model: 4096 if model == "haiku" else 1024,
)
def _formula(m: ItemMeasure, u: CacheMix) -> float:
"""The closed form the record must agree with."""
billed = u.read + u.write_5m + u.uncached
k = billed / m.sent_tokens
warm = min(1.0, (u.read / billed) / (m.sent_shared_prefix / m.sent_tokens))
w = u.write_5m / (u.uncached + u.write_5m)
p_miss = w * WRITE + (1 - w) * IN
return (
k * m.removed_old * (warm * READ + (1 - warm) * p_miss)
+ k * m.removed_new * p_miss
- k * m.bust * warm * (p_miss - READ)
)
WARM = CacheMix(uncached=2_000, read=90_000, write_5m=8_000)
MEASURE = ItemMeasure(
sent_tokens=95_000, removed_old=50_000, removed_new=10_000, sent_shared_prefix=88_000, bust=0
)
def test_worked_example():
record = compute_savings_record(MEASURE, WARM, model="sonnet", provider="anthropic")
assert record.k == pytest.approx(1.0526, abs=1e-4)
assert record.warm == pytest.approx(0.97159, abs=1e-5)
assert record.write_share == pytest.approx(0.8)
assert record.saved_usd == pytest.approx(0.058618, abs=1e-6)
assert record.saved_usd == pytest.approx(_formula(MEASURE, WARM), abs=1e-9)
assert record.saved_usd == pytest.approx(
record.counterfactual_input_cost_usd - record.actual_input_cost_usd
)
assert record.saved_tokens == pytest.approx(100_000 / 95_000 * 60_000)
assert not record.estimated
def test_a_cache_bust_headroom_caused_is_charged_against_it():
busted = dataclasses.replace(MEASURE, bust=5_000)
record = compute_savings_record(busted, WARM, model="sonnet", provider="anthropic")
assert record.saved_usd == pytest.approx(_formula(busted, WARM), abs=1e-9)
assert record.saved_usd == pytest.approx(0.058618 - 0.016875, abs=1e-6)
def test_a_bust_never_claims_more_reads_than_the_request_missed():
"""Item-level P_S undercounts a token-level hit, so hit > sigma clips warm
to 1; the bust charge must still not drive the counterfactual negative."""
measure = ItemMeasure(
sent_tokens=10_000, removed_old=0, removed_new=0, sent_shared_prefix=2_000, bust=8_000
)
usage = CacheMix(uncached=500, read=9_000, write_5m=500)
record = compute_savings_record(measure, usage, model="sonnet", provider="anthropic")
assert record.counterfactual_input_cost_usd >= 0
# At most the 1,000 missed tokens could have been reads instead.
assert record.saved_usd == pytest.approx(-(500 * (IN - READ) + 500 * (WRITE - READ)))
def test_an_old_item_headroom_grew_inherits_no_reads_below_the_cache_minimum():
"""JerrettDavis's #4236 case: a 1,000-token item sent as 10,000 tokens on
successive requests, prompt cached. The 1,000-token counterfactual is
under the 1,024-token minimum, so none of it could have been a read."""
measure = ItemMeasure(
sent_tokens=10_000, removed_old=-9_000, removed_new=0, sent_shared_prefix=10_000, bust=0
)
usage = CacheMix(read=9_000, uncached=1_000)
for provider in ("openai", "anthropic"):
record = compute_savings_record(measure, usage, model="gpt", provider=provider)
assert record.actual_input_cost_usd == pytest.approx(0.0057)
assert record.counterfactual_input_cost_usd == pytest.approx(1_000 * IN) # $0.003
assert record.saved_usd == pytest.approx(-0.0027)
def test_summed_iteration_usage_does_not_calibrate_k():
"""Usage summed over server-side iterations describes more prompts than
the one sent body, so it must not scale the removal."""
summed = CacheMix(read=270_000, uncached=6_000, write_5m=24_000)
record = compute_savings_record(
MEASURE, summed, model="sonnet", provider="anthropic", estimated=True
)
assert record.k == 1.0
assert record.estimated
assert record.saved_tokens == 60_000
def test_a_cold_cache_prices_every_removed_token_as_a_miss():
cold = CacheMix(uncached=10_000, read=0, write_5m=90_000)
record = compute_savings_record(MEASURE, cold, model="sonnet", provider="anthropic")
assert record.warm == 0.0
p_miss = 0.9 * WRITE + 0.1 * IN
assert record.saved_usd == pytest.approx(100_000 / 95_000 * 60_000 * p_miss)
def test_one_hour_writes_price_at_their_own_rate():
usage = CacheMix(uncached=2_000, read=90_000, write_1h=8_000)
record = compute_savings_record(MEASURE, usage, model="sonnet", provider="anthropic")
k, warm = 100_000 / 95_000, record.warm
p_miss = 0.8 * WRITE_1H + 0.2 * IN
expected = k * 50_000 * (warm * READ + (1 - warm) * p_miss) + k * 10_000 * p_miss
assert record.saved_usd == pytest.approx(expected)
def test_an_inferred_openai_write_is_the_uncached_slice_not_a_write():
usage = CacheMix(read=90_000, write_total=10_000, uncached=10_000, inferred_write=True)
record = compute_savings_record(MEASURE, usage, model="gpt", provider="openai")
assert record.k == pytest.approx(100_000 / 95_000)
assert record.write_share == 0.0
def test_openai_reads_come_in_128_token_blocks():
measure = ItemMeasure(
sent_tokens=10_000, removed_old=1_000, removed_new=0, sent_shared_prefix=10_000, bust=0
)
usage = CacheMix(read=9_984, uncached=16) # 78 blocks
record = compute_savings_record(measure, usage, model="gpt", provider="openai")
# 1,000 * k removed reads would end mid-block: only whole blocks count.
total_read = 9_984 + 1_000 * record.k
reads = (total_read // 128) * 128 - 9_984
assert record.saved_usd == pytest.approx(reads * READ + (1_000 * record.k - reads) * IN)
def test_a_counterfactual_prefix_below_the_models_cache_minimum_is_never_warm():
# A 3,000-token old prefix: cacheable on Sonnet (1,024), not on Haiku (4,096).
measure = ItemMeasure(
sent_tokens=10_000, removed_old=1_000, removed_new=0, sent_shared_prefix=2_000, bust=0
)
usage = CacheMix(uncached=5_000, read=5_000)
assert compute_savings_record(measure, usage, model="sonnet").warm > 0
assert compute_savings_record(measure, usage, model="haiku").warm == 0.0
def test_the_counterfactual_is_priced_at_its_own_context_tier():
measure = ItemMeasure(
sent_tokens=150_000, removed_old=0, removed_new=100_000, sent_shared_prefix=0, bust=0
)
record = compute_savings_record(measure, CacheMix(uncached=150_000), model="sonnet")
assert record.actual_input_cost_usd == pytest.approx(150_000 * IN)
assert record.counterfactual_input_cost_usd == pytest.approx(250_000 * 2 * IN)
def test_missing_usage_prices_at_list_per_tier_and_says_so():
measure = dataclasses.replace(MEASURE, sent_tokens=150_000, removed_new=50_000)
record = compute_savings_record(measure, None, model="sonnet")
assert record.estimated
assert record.actual_input_cost_usd == pytest.approx(150_000 * IN)
assert record.counterfactual_input_cost_usd == pytest.approx(250_000 * 2 * IN)
def test_an_unpriced_model_reports_tokens_but_no_dollars():
record = compute_savings_record(MEASURE, WARM, model="unpriced")
assert not record.priced
assert record.saved_usd == 0.0
assert record.removed_tokens == 60_000
def test_items_split_into_old_and_new_against_the_previous_request():
def item(name: str, original: int, sent: int, sent_name: str | None = None) -> ItemCounts:
return ItemCounts(name, original, sent_name or name + "'", sent)
first = [item("tools", 1_000, 200), item("sys", 500, 500, "sys"), item("u1", 300, 300, "u1")]
second = [
item("tools", 1_000, 200),
item("sys", 500, 500, "sys"),
# Headroom compressed this item differently this time: its sent copy
# changed, so the provider missed it although the client resent it.
item("u1", 300, 250, "u1-recompressed"),
item("t1", 4_000, 1_000),
# An item Headroom grew counts against the removal.
item("t2", 100, 160),
]
assert measure_items(first, None) == ItemMeasure(
sent_tokens=1_000, removed_old=0, removed_new=800, sent_shared_prefix=0, bust=0
)
assert measure_items(second, first) == ItemMeasure(
sent_tokens=2_110,
removed_old=800 + 50,
removed_new=3_000 - 60,
sent_shared_prefix=700,
bust=250,
grown_new=60,
)
def test_billed_buckets_price_at_billing_rates(monkeypatch):
"""A catalog row without a cache-read rate bills reads via litellm's
calculator (``for_billing``); only the changed tokens use the estimate."""
billing = dataclasses.replace(BASE, read=5e-7)
def rates(model, *, long_context=False, provider=None, for_billing=False):
return billing if for_billing else BASE
monkeypatch.setattr(counterfactual, "resolve_rates", rates)
record = compute_savings_record(MEASURE, WARM, model="sonnet", provider="anthropic")
assert record.actual_input_cost_usd == pytest.approx(2_000 * IN + 90_000 * 5e-7 + 8_000 * WRITE)
assert record.saved_usd == pytest.approx(0.058618, abs=1e-6)
def _fully_cached_original_of(original: int, *, model: str, provider: str):
"""A 10,000-token sent body read wholly from cache, whose original was
``original`` tokens: the counterfactual is that original alone."""
measure = ItemMeasure(
sent_tokens=10_000,
removed_old=original - 10_000,
removed_new=0,
sent_shared_prefix=10_000,
bust=0,
)
return compute_savings_record(measure, CacheMix(read=10_000), model=model, provider=provider)
@pytest.mark.parametrize(
("original", "reads"),
[(1_023, 0), (1_024, 1_024), (1_151, 1_024), (1_152, 1_152), (2_000, 1_920)],
)
def test_an_openai_counterfactual_reads_whole_blocks_above_its_own_minimum(original, reads):
record = _fully_cached_original_of(original, model="gpt", provider="openai")
expected = reads * READ + (original - reads) * IN
assert record.counterfactual_input_cost_usd == pytest.approx(expected)
@pytest.mark.parametrize(("original", "reads"), [(4_095, 0), (4_096, 4_096), (4_097, 4_097)])
def test_an_anthropic_counterfactual_obeys_its_models_cache_minimum(original, reads):
record = _fully_cached_original_of(original, model="haiku", provider="anthropic")
expected = reads * READ + (original - reads) * IN
assert record.counterfactual_input_cost_usd == pytest.approx(expected)
def test_a_counterfactual_too_small_to_cache_pays_no_write_premium():
"""Below its minimum the counterfactual can be neither read nor written."""
measure = ItemMeasure(
sent_tokens=10_000, removed_old=-9_000, removed_new=0, sent_shared_prefix=10_000, bust=0
)
usage = CacheMix(read=9_000, write_5m=1_000)
record = compute_savings_record(measure, usage, model="sonnet", provider="anthropic")
assert record.counterfactual_input_cost_usd == pytest.approx(1_000 * IN)
def test_a_grown_old_item_takes_its_reads_out_even_with_new_items():
"""The new items keep the counterfactual above the minimum overall; its
1,000-token old prefix is still too small to have been read."""
measure = ItemMeasure(
sent_tokens=30_000,
removed_old=-9_000,
removed_new=20_000,
sent_shared_prefix=10_000,
bust=0,
)
usage = CacheMix(read=9_000, uncached=21_000)
record = compute_savings_record(measure, usage, model="gpt", provider="openai")
assert record.counterfactual_input_cost_usd == pytest.approx(41_000 * IN)
def test_a_catalog_minimum_the_provider_disproved_is_not_applied():
"""Haiku's catalog says 4,096, but the provider read 3,000: keeping the
counterfactual's 1,500 reads is the conservative price."""
measure = ItemMeasure(
sent_tokens=10_000, removed_old=-5_000, removed_new=0, sent_shared_prefix=10_000, bust=0
)
usage = CacheMix(read=3_000, uncached=7_000)
record = compute_savings_record(measure, usage, model="haiku", provider="anthropic")
assert record.counterfactual_input_cost_usd == pytest.approx(1_500 * READ + 3_500 * IN)
def test_nothing_removed_saves_nothing_whatever_the_read_alignment():
measure = ItemMeasure(
sent_tokens=10_000, removed_old=0, removed_new=0, sent_shared_prefix=2_000, bust=0
)
for provider, usage in (
("openai", CacheMix(read=9_000, uncached=1_000)),
("anthropic", CacheMix(read=900, write_5m=9_100)),
):
record = compute_savings_record(measure, usage, model="haiku", provider=provider)
assert record.saved_usd == pytest.approx(0.0, abs=1e-12)
def test_a_shrunk_bucket_is_never_priced_below_zero(monkeypatch):
"""Reads a grown item gave up price at the billing rate they were billed
at, not at the estimate rate, so the counterfactual cannot go negative."""
billing = dataclasses.replace(BASE, read=1.5e-6)
def rates(model, *, long_context=False, provider=None, for_billing=False):
return billing if for_billing else BASE
monkeypatch.setattr(counterfactual, "resolve_rates", rates)
measure = ItemMeasure(
sent_tokens=10_000, removed_old=-9_000, removed_new=0, sent_shared_prefix=10_000, bust=0
)
record = compute_savings_record(
measure, CacheMix(read=9_000, uncached=1_000), model="gpt", provider="openai"
)
assert record.counterfactual_input_cost_usd == pytest.approx(1_000 * IN)
def test_a_disproved_minimum_still_writes_nothing_smaller_than_was_cached():
"""Haiku's catalog minimum is disproved by a 2,000-token read; a 500-token
counterfactual is below anything shown cacheable, so it writes nothing."""
measure = ItemMeasure(
sent_tokens=3_000, removed_old=-2_500, removed_new=0, sent_shared_prefix=3_000, bust=0
)
usage = CacheMix(read=2_000, write_5m=1_000)
record = compute_savings_record(measure, usage, model="haiku", provider="anthropic")
reads = 2_000 + -2_500 * (2 / 3)
assert record.counterfactual_input_cost_usd == pytest.approx(reads * READ + (500 - reads) * IN)
def test_a_lost_baseline_prices_the_new_removal_at_the_read_floor():
"""With the conversation's baseline evicted, the new items may be old ones
the provider had cached. Pricing them as new, at the miss rate, booked
$0.2236 here; the read rate, the floor of any input token, books $0.0152.
The 1,000 tokens Headroom added to a new item still cost the miss rate."""
lost = ItemMeasure(
sent_tokens=95_000,
removed_old=0,
removed_new=60_000 - 1_000,
sent_shared_prefix=0,
bust=0,
aligned=False,
baseline_lost=True,
grown_new=1_000,
)
k, p_miss = 100_000 / 95_000, 0.8 * WRITE + 0.2 * IN
as_new = compute_savings_record(
dataclasses.replace(lost, baseline_lost=False), WARM, model="sonnet", provider="anthropic"
)
assert as_new.saved_usd == pytest.approx(k * 59_000 * p_miss) # $0.2236
record = compute_savings_record(lost, WARM, model="sonnet", provider="anthropic")
assert record.saved_usd == pytest.approx(k * 60_000 * READ - k * 1_000 * p_miss) # $0.0152
assert record.saved_usd == pytest.approx(0.015158, abs=1e-6)
assert record.estimated
# Without usage too: list price for the rest, the read rate for the removal.
unbilled = compute_savings_record(lost, None, model="sonnet", provider="anthropic")
assert unbilled.saved_usd == pytest.approx(60_000 * READ - 1_000 * IN)