feat(pricing): real per-model cost catalog; fix non-Claude mispricing

Replaces the hardcoded 5-model Claude price table (extract.ts) with a curated, websearch-sourced catalog of 63 models across Anthropic/OpenAI/Google/Chinese/other vendors, behind a small lookup/compute module (src/pricing/catalog.ts). Non-Claude models now price off their own rows instead of defaulting to Claude Sonnet. Adds provider/ prefix normalization, native-cost passthrough, and null+warn on unknown models. LiteLLM ~2900-model catalog kept as a dev-only refresh base (tools/pricing/, gitignored), not bundled.

.gitignore

@@ -39,5 +39,9 @@ context-mode-guidance-*/
 .claude-worktrees/
 .vibetree/
 .cw/
+# Dev-only pricing source: the full litellm catalog (~1.5MB) is the manual
+# refresh base for src/pricing/sources/*.json — never bundled into the plugin.
+# Refresh: see tools/pricing/litellm-NOTES.md. Keep NOTES tracked, ignore the blob.
+tools/pricing/litellm-catalog.json
 /.cocoindex_code/
 /.kilo/

src/pricing/catalog.ts

@@ -0,0 +1,203 @@
+/**
+ * Pricing catalog — single source of truth for per-model USD cost.
+ *
+ * Deep module, tiny interface. At load it merges the 5 curated vendor JSONs
+ * (src/pricing/sources/{anthropic,openai,google,chinese,others}.json) into one
+ * Map<modelId, Price> in per-Mtok units, then exposes three pure-ish functions:
+ *
+ *   lookupPrice(modelId)            → Price | null
+ *   computeCostUsd(modelId, tokens) → number  | null
+ *   nativeOrComputed(id, t, native) → number  | null
+ *
+ * WHY THIS EXISTS — the bug it kills:
+ * The old table in src/session/extract.ts hardcoded ~5 Claude rows plus a
+ * `default` row, and any unmatched id (every OpenAI / Gemini / Qwen / DeepSeek
+ * / Grok model) silently inherited Claude-Sonnet pricing. Non-Claude turns were
+ * therefore mispriced. Here each model is priced from ITS OWN curated row, and
+ * an unknown id resolves to `null` (no price) instead of a wrong Claude rate.
+ *
+ * The large litellm catalog (~1.5MB, ~2900 models) is NOT bundled — it lives at
+ * tools/pricing/litellm-catalog.json as the dev-only refresh base for these
+ * curated JSONs. See tools/pricing/litellm-NOTES.md for the refresh recipe.
+ *
+ * The 5 curated JSONs are small (~25KB total) and esbuild inlines them into the
+ * hook/server bundles at build time (no runtime fs read, no external file).
+ */
+
+import anthropic from "./sources/anthropic.json" with { type: "json" };
+import openai from "./sources/openai.json" with { type: "json" };
+import google from "./sources/google.json" with { type: "json" };
+import chinese from "./sources/chinese.json" with { type: "json" };
+import others from "./sources/others.json" with { type: "json" };
+
+/** Per-Mtok price for one model. Any of the four rates may be null ("unknown"). */
+export interface Price {
+  input_per_mtok: number | null;
+  output_per_mtok: number | null;
+  cache_read_per_mtok: number | null;
+  cache_write_per_mtok: number | null;
+}
+
+/** Token counts for one turn. All fields optional; absent ⇒ treated as 0. */
+export interface TokenCounts {
+  input_tokens?: number;
+  output_tokens?: number;
+  cache_read_tokens?: number;
+  cache_creation_tokens?: number;
+}
+
+/** Raw shape of a curated source row (carries provenance fields we drop). */
+interface RawRow {
+  input_per_mtok: number | null;
+  output_per_mtok: number | null;
+  cache_read_per_mtok: number | null;
+  cache_write_per_mtok: number | null;
+  [extra: string]: unknown;
+}
+
+/**
+ * Merge the 5 vendor JSONs into one Map. A row with a null *input* price is
+ * unusable for cost (the primary bucket has no rate) and is dropped at load so
+ * lookupPrice returns null for it — matching "null-priced entries → no price".
+ * Curated ids are globally unique across the five files (verified), so merge
+ * order is irrelevant; later files would otherwise win on collision.
+ */
+function buildCatalog(): Map<string, Price> {
+  const map = new Map<string, Price>();
+  const sources: Record<string, RawRow>[] = [
+    anthropic as Record<string, RawRow>,
+    openai as Record<string, RawRow>,
+    google as Record<string, RawRow>,
+    chinese as Record<string, RawRow>,
+    others as Record<string, RawRow>,
+  ];
+  for (const src of sources) {
+    for (const id of Object.keys(src)) {
+      const row = src[id];
+      if (row == null || typeof row !== "object") continue;
+      // No input rate ⇒ no usable price for this model.
+      if (typeof row.input_per_mtok !== "number") continue;
+      map.set(id, {
+        input_per_mtok: row.input_per_mtok,
+        output_per_mtok: typeof row.output_per_mtok === "number" ? row.output_per_mtok : null,
+        cache_read_per_mtok:
+          typeof row.cache_read_per_mtok === "number" ? row.cache_read_per_mtok : null,
+        cache_write_per_mtok:
+          typeof row.cache_write_per_mtok === "number" ? row.cache_write_per_mtok : null,
+      });
+    }
+  }
+  return map;
+}
+
+const CATALOG: Map<string, Price> = buildCatalog();
+
+/**
+ * Strip a single leading `provider/` segment, char-algorithmically (NO regex).
+ * Walks to the first '/'; everything after it is the bare model id. Only the
+ * FIRST segment is stripped — `openai/gpt-5` → `gpt-5`, `a/b/c` → `b/c` — so a
+ * model id that legitimately contains a slash keeps its remaining segments.
+ * Returns null when there is no '/' (caller already tried the raw form).
+ */
+function stripProviderPrefix(id: string): string | null {
+  for (let i = 0; i < id.length; i++) {
+    if (id.charCodeAt(i) === 47 /* '/' */) {
+      // Guard against a leading or trailing slash producing an empty segment.
+      if (i === 0 || i === id.length - 1) return null;
+      return id.slice(i + 1);
+    }
+  }
+  return null;
+}
+
+/** trim + lowercase, char-safe (String.prototype.trim/toLowerCase, no regex). */
+function normalize(id: string): string {
+  return id.trim().toLowerCase();
+}
+
+/**
+ * Resolve a model id to its curated Price, or null on miss.
+ * Strategy: exact, then normalized (trim+lowercase), then provider-stripped
+ * normalized. Misses return null so the caller can decide (warn / fall back).
+ */
+export function lookupPrice(modelId: string): Price | null {
+  if (typeof modelId !== "string" || modelId.length === 0) return null;
+
+  // 1. Exact — fastest path, covers ids already in canonical form.
+  const exact = CATALOG.get(modelId);
+  if (exact) return exact;
+
+  // 2. Normalized — trimmed + lowercased.
+  const norm = normalize(modelId);
+  const byNorm = CATALOG.get(norm);
+  if (byNorm) return byNorm;
+
+  // 3. Provider-stripped (pi / openclaw / opencode report `provider/model`).
+  const bare = stripProviderPrefix(norm);
+  if (bare) {
+    const byBare = CATALOG.get(bare);
+    if (byBare) return byBare;
+  }
+
+  return null;
+}
+
+/** Price one token bucket. A null bucket rate falls back to the input rate. */
+function bucketCost(tokens: number, rate: number | null, inputRate: number): number {
+  if (tokens <= 0) return 0;
+  const effective = typeof rate === "number" ? rate : inputRate;
+  return tokens * effective;
+}
+
+/**
+ * Σ tokens × per-Mtok price / 1e6 over the four buckets. Returns null when no
+ * price is found OR every token count is zero/absent (so dashboards never show
+ * a misleading "$0.00 for nothing" row). On a price miss, warns exactly once
+ * with the unmatched id so the curated catalog can be extended.
+ *
+ * Bucket → price mapping:
+ *   input_tokens          → input_per_mtok
+ *   output_tokens         → output_per_mtok  (null ⇒ input rate)
+ *   cache_read_tokens     → cache_read_per_mtok  (null ⇒ input rate)
+ *   cache_creation_tokens → cache_write_per_mtok (null ⇒ input rate)
+ */
+export function computeCostUsd(modelId: string, t: TokenCounts): number | null {
+  const input = typeof t.input_tokens === "number" ? t.input_tokens : 0;
+  const output = typeof t.output_tokens === "number" ? t.output_tokens : 0;
+  const cacheRead = typeof t.cache_read_tokens === "number" ? t.cache_read_tokens : 0;
+  const cacheCreate =
+    typeof t.cache_creation_tokens === "number" ? t.cache_creation_tokens : 0;
+
+  // All buckets empty ⇒ nothing to price, regardless of model.
+  if (input <= 0 && output <= 0 && cacheRead <= 0 && cacheCreate <= 0) return null;
+
+  const price = lookupPrice(modelId);
+  if (!price || typeof price.input_per_mtok !== "number") {
+    // Unknown model — emit one line so the id can be added to the catalog.
+    console.warn(`[pricing] no curated price for model id: ${modelId}`);
+    return null;
+  }
+
+  const inputRate = price.input_per_mtok;
+  const microDollars =
+    bucketCost(input, inputRate, inputRate) +
+    bucketCost(output, price.output_per_mtok, inputRate) +
+    bucketCost(cacheRead, price.cache_read_per_mtok, inputRate) +
+    bucketCost(cacheCreate, price.cache_write_per_mtok, inputRate);
+
+  return microDollars / 1_000_000;
+}
+
+/**
+ * Prefer a provider-supplied native cost when present, else compute from the
+ * catalog. A native cost of exactly 0 is a real value (free tier) and passes
+ * through — only null/undefined defers to computeCostUsd.
+ */
+export function nativeOrComputed(
+  modelId: string,
+  t: TokenCounts,
+  nativeCostUsd?: number | null,
+): number | null {
+  if (typeof nativeCostUsd === "number") return nativeCostUsd;
+  return computeCostUsd(modelId, t);
+}

src/pricing/sources/anthropic.json

@@ -0,0 +1,82 @@
+{
+  "claude-opus-4-8": {
+    "input_per_mtok": 5.00,
+    "output_per_mtok": 25.00,
+    "cache_read_per_mtok": 0.50,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-opus-4-7": {
+    "input_per_mtok": 5.00,
+    "output_per_mtok": 25.00,
+    "cache_read_per_mtok": 0.50,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-opus-4-6": {
+    "input_per_mtok": 5.00,
+    "output_per_mtok": 25.00,
+    "cache_read_per_mtok": 0.50,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-opus-4-5": {
+    "input_per_mtok": 5.00,
+    "output_per_mtok": 25.00,
+    "cache_read_per_mtok": 0.50,
+    "cache_write_per_mtok": 6.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-sonnet-4-6": {
+    "input_per_mtok": 3.00,
+    "output_per_mtok": 15.00,
+    "cache_read_per_mtok": 0.30,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-sonnet-4-5": {
+    "input_per_mtok": 3.00,
+    "output_per_mtok": 15.00,
+    "cache_read_per_mtok": 0.30,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-haiku-4-5": {
+    "input_per_mtok": 1.00,
+    "output_per_mtok": 5.00,
+    "cache_read_per_mtok": 0.10,
+    "cache_write_per_mtok": 1.25,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-3-7-sonnet": {
+    "input_per_mtok": 3.00,
+    "output_per_mtok": 15.00,
+    "cache_read_per_mtok": 0.30,
+    "cache_write_per_mtok": 3.75,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "claude-3-5-haiku": {
+    "input_per_mtok": 0.80,
+    "output_per_mtok": 4.00,
+    "cache_read_per_mtok": 0.08,
+    "cache_write_per_mtok": 1.00,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  },
+  "claude-fable-5": {
+    "input_per_mtok": 10.00,
+    "output_per_mtok": 50.00,
+    "cache_read_per_mtok": 1.00,
+    "cache_write_per_mtok": 12.50,
+    "source": "https://platform.claude.com/docs/en/about-claude/pricing",
+    "as_of": "2026-06"
+  }
+}

src/pricing/sources/chinese.json

@@ -0,0 +1,155 @@
+{
+  "qwen3-coder": {
+    "input_per_mtok": 1.0,
+    "output_per_mtok": 5.0,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+    "as_of": "2026-06",
+    "_note": "Billing id maps to qwen3-coder-plus. Tiered by input length; headline is the 0-32K tier. Higher tiers: 32K-128K in 1.8/out 9, 128K-256K in 3/out 15, 256K-1M in 6/out 60. Cross-checked vs LiteLLM tiered_pricing."
+  },
+  "qwen-max": {
+    "input_per_mtok": 1.6,
+    "output_per_mtok": 6.4,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+    "as_of": "2026-06",
+    "_note": "Alibaba DashScope international (USD-denominated). Cross-checked vs LiteLLM dashscope/qwen-max."
+  },
+  "qwen-plus": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 1.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+    "as_of": "2026-06",
+    "_note": "Alibaba DashScope international (USD). Cross-checked vs LiteLLM dashscope/qwen-plus."
+  },
+  "qwen-turbo": {
+    "input_per_mtok": 0.05,
+    "output_per_mtok": 0.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+    "as_of": "2026-06",
+    "_note": "Alibaba DashScope international (USD). Cross-checked vs LiteLLM dashscope/qwen-turbo."
+  },
+  "qwen3-max": {
+    "input_per_mtok": 1.2,
+    "output_per_mtok": 6.0,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+    "as_of": "2026-06",
+    "_note": "Tiered by input length; headline is the 0-32K tier. Higher tiers: 32K-128K in 2.4/out 12, 128K-252K in 3/out 15. Cross-checked vs LiteLLM dashscope/qwen3-max tiered_pricing."
+  },
+  "kimi-k2": {
+    "input_per_mtok": 0.6,
+    "output_per_mtok": 2.5,
+    "cache_read_per_mtok": 0.15,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing/chat",
+    "as_of": "2026-06",
+    "_note": "Moonshot international platform (USD). Resolves to kimi-k2-0905-preview. Cache-hit input 0.15. Cross-checked vs LiteLLM moonshot/kimi-k2-0905-preview."
+  },
+  "kimi-k2-turbo": {
+    "input_per_mtok": 1.15,
+    "output_per_mtok": 8.0,
+    "cache_read_per_mtok": 0.15,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing/chat",
+    "as_of": "2026-06",
+    "_note": "Moonshot international (USD). Resolves to kimi-k2-turbo-preview. Cache-hit input 0.15. Cross-checked vs LiteLLM moonshot/kimi-k2-turbo-preview."
+  },
+  "moonshot-v1-8k": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 2.0,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-8k."
+  },
+  "moonshot-v1-32k": {
+    "input_per_mtok": 1.0,
+    "output_per_mtok": 3.0,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-32k."
+  },
+  "moonshot-v1-128k": {
+    "input_per_mtok": 2.0,
+    "output_per_mtok": 5.0,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://platform.moonshot.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Moonshot international (USD). Cross-checked vs LiteLLM moonshot/moonshot-v1-128k."
+  },
+  "deepseek-v3": {
+    "input_per_mtok": 0.27,
+    "output_per_mtok": 1.1,
+    "cache_read_per_mtok": 0.07,
+    "cache_write_per_mtok": 0.0,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy DeepSeek-V3 era pricing. No longer shown on the official current pricing page (which now lists deepseek-v4-flash/pro). Value from LiteLLM deepseek/deepseek-v3, last sourced from the official page. cache-hit input 0.07."
+  },
+  "deepseek-r1": {
+    "input_per_mtok": 0.55,
+    "output_per_mtok": 2.19,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy DeepSeek-R1 pricing. Not on the official current pricing page (now deepseek-v4-flash/pro). Value from LiteLLM deepseek/deepseek-r1."
+  },
+  "deepseek-chat": {
+    "input_per_mtok": 0.14,
+    "output_per_mtok": 0.28,
+    "cache_read_per_mtok": 0.0028,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing",
+    "as_of": "2026-06",
+    "_note": "Official page (fetched 2026-06): deepseek-chat is deprecated 2026/07/24 and now maps to deepseek-v4-flash non-thinking mode. Live billing = v4-flash: input(cache-miss) 0.14, output 0.28, input(cache-hit) 0.0028."
+  },
+  "deepseek-reasoner": {
+    "input_per_mtok": 0.14,
+    "output_per_mtok": 0.28,
+    "cache_read_per_mtok": 0.0028,
+    "cache_write_per_mtok": null,
+    "source": "https://api-docs.deepseek.com/quick_start/pricing",
+    "as_of": "2026-06",
+    "_note": "Official page (fetched 2026-06): deepseek-reasoner is deprecated 2026/07/24 and now maps to deepseek-v4-flash thinking mode. Live billing = v4-flash: input(cache-miss) 0.14, output 0.28, input(cache-hit) 0.0028."
+  },
+  "glm-4.6": {
+    "input_per_mtok": 0.6,
+    "output_per_mtok": 2.2,
+    "cache_read_per_mtok": 0.11,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.z.ai/guides/overview/pricing",
+    "as_of": "2026-06",
+    "_note": "Z.AI international (USD). Cached-input 0.11. Cache-input storage currently limited-time free. Cross-checked vs LiteLLM zai/glm-4.6."
+  },
+  "glm-4-plus": {
+    "input_per_mtok": null,
+    "output_per_mtok": null,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": null,
+    "as_of": "2026-06",
+    "_note": "No authoritative current price found this session. glm-4-plus is a legacy Zhipu/BigModel-native model not listed on the Z.AI international catalog; open.bigmodel.cn pricing is SPA-rendered (no static price) and BigModel English docs returned HTTP 552/404. Not in LiteLLM. Left null per anti-hallucination rule."
+  },
+  "glm-4-air": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 1.1,
+    "cache_read_per_mtok": 0.03,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.z.ai/guides/overview/pricing",
+    "as_of": "2026-06",
+    "_note": "Mission billing id glm-4-air; current Z.AI international catalog lists this as GLM-4.5-Air (USD): input 0.2, cached-input 0.03, output 1.1. Cross-checked vs LiteLLM zai/glm-4.5-air (in 0.2/out 1.1)."
+  }
+}

src/pricing/sources/google.json

@@ -0,0 +1,74 @@
+{
+  "gemini-2.5-pro": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10.0,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Tiered: prompts >200k tokens cost input $2.50, output $15.00, cache_read $0.25 per Mtok. Cache storage $4.50/Mtok/hr."
+  },
+  "gemini-2.5-flash": {
+    "input_per_mtok": 0.3,
+    "output_per_mtok": 2.5,
+    "cache_read_per_mtok": 0.03,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Text/image/video input; audio input $1.00. Output includes thinking tokens. Cache storage $1.00/Mtok/hr. No >200k context tier."
+  },
+  "gemini-2.5-flash-lite": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.01,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Text/image/video input; audio input $0.30. Cache storage $1.00/Mtok/hr. No >200k context tier."
+  },
+  "gemini-2.0-flash": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Text/image/video input; audio input $0.70, audio cache_read $0.175. Cache storage $1.00/Mtok/hr. No >200k context tier."
+  },
+  "gemini-2.0-flash-lite": {
+    "input_per_mtok": 0.075,
+    "output_per_mtok": 0.3,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "DEPRECATED — shut down June 1, 2026. Context caching not available for this model. Prices retained for historical billing."
+  },
+  "gemini-2.0-pro": {
+    "input_per_mtok": null,
+    "output_per_mtok": null,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "No GA billing id. Only shipped as gemini-2.0-pro-exp (experimental, free of charge); never had paid pricing. Not listed on official pricing page or LiteLLM catalog."
+  },
+  "gemini-3-pro-preview": {
+    "input_per_mtok": 2.0,
+    "output_per_mtok": 12.0,
+    "cache_read_per_mtok": 0.2,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Current 3.x flagship (page labels it Gemini 3.1 Pro Preview). Tiered: prompts >200k cost input $4.00, output $18.00, cache_read $0.40 per Mtok. Cache storage $4.50/Mtok/hr."
+  },
+  "gemini-3-flash-preview": {
+    "input_per_mtok": 0.5,
+    "output_per_mtok": 3.0,
+    "cache_read_per_mtok": 0.05,
+    "cache_write_per_mtok": null,
+    "source": "https://ai.google.dev/gemini-api/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Text/image/video input; audio input $1.00, audio cache_read $0.10. Cache storage $1.00/Mtok/hr. No >200k context tier on standard."
+  }
+}

src/pricing/sources/openai.json

@@ -0,0 +1,106 @@
+{
+  "gpt-5": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-5-mini": {
+    "input_per_mtok": 0.25,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-5-nano": {
+    "input_per_mtok": 0.05,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.005,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-5-codex": {
+    "input_per_mtok": 1.25,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 0.125,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-4.1": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 8,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-4.1-mini": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 1.6,
+    "cache_read_per_mtok": 0.1,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-4.1-nano": {
+    "input_per_mtok": 0.1,
+    "output_per_mtok": 0.4,
+    "cache_read_per_mtok": 0.025,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-4o": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": 1.25,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "gpt-4o-mini": {
+    "input_per_mtok": 0.15,
+    "output_per_mtok": 0.6,
+    "cache_read_per_mtok": 0.075,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "o3": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 8,
+    "cache_read_per_mtok": 0.5,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "o4-mini": {
+    "input_per_mtok": 1.1,
+    "output_per_mtok": 4.4,
+    "cache_read_per_mtok": 0.275,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "o3-mini": {
+    "input_per_mtok": 1.1,
+    "output_per_mtok": 4.4,
+    "cache_read_per_mtok": 0.55,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  },
+  "codex-mini-latest": {
+    "input_per_mtok": 1.5,
+    "output_per_mtok": 6,
+    "cache_read_per_mtok": 0.375,
+    "cache_write_per_mtok": null,
+    "source": "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json",
+    "as_of": "2026-06"
+  }
+}

src/pricing/sources/others.json

@@ -0,0 +1,137 @@
+{
+  "grok-4": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy id; xAI public pricing page now lists newer Grok versions (4.3/4.20/build-0.1). Rate verified via LiteLLM xai/ catalog (BerriAI/litellm) which sources directly from xAI."
+  },
+  "grok-3": {
+    "input_per_mtok": 3,
+    "output_per_mtok": 15,
+    "cache_read_per_mtok": 0.75,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy id; rate from LiteLLM xai/grok-3 catalog (sourced from xAI)."
+  },
+  "grok-code-fast-1": {
+    "input_per_mtok": 0.2,
+    "output_per_mtok": 1.5,
+    "cache_read_per_mtok": 0.02,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy id; rate from LiteLLM xai/grok-code-fast-1 catalog (sourced from xAI)."
+  },
+  "grok-2": {
+    "input_per_mtok": 2,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://docs.x.ai/docs/pricing",
+    "as_of": "2026-06",
+    "_note": "Legacy id (grok-2-1212); rate from LiteLLM xai/grok-2 catalog (sourced from xAI)."
+  },
+  "mistral-large-latest": {
+    "input_per_mtok": 0.5,
+    "output_per_mtok": 1.5,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "mistral-large-latest = Mistral Large 3."
+  },
+  "codestral-latest": {
+    "input_per_mtok": 0.3,
+    "output_per_mtok": 0.9,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Codestral 2; price cross-confirmed via Google Vertex partner pricing ($0.30/$0.90)."
+  },
+  "devstral": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Maps to devstral-medium-latest (Mistral hosted API). devstral-small is $0.10/$0.30."
+  },
+  "mistral-medium": {
+    "input_per_mtok": 0.4,
+    "output_per_mtok": 2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://mistral.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Mistral Medium 3; cross-confirmed via Google Vertex partner pricing ($0.40/$2.00)."
+  },
+  "llama-4-maverick": {
+    "input_per_mtok": 0.27,
+    "output_per_mtok": 0.85,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Host-dependent. Priced via Together AI serverless. Other hosts differ (e.g. AWS Bedrock, Vertex $0.35/$1.15)."
+  },
+  "llama-4-scout": {
+    "input_per_mtok": 0.08,
+    "output_per_mtok": 0.3,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Host-dependent. Priced via Together AI serverless. Vertex lists $0.25/$0.70."
+  },
+  "llama-3.3-70b": {
+    "input_per_mtok": 0.88,
+    "output_per_mtok": 0.88,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://www.together.ai/pricing",
+    "as_of": "2026-06",
+    "_note": "Host-dependent. Together Llama-3.3-70B-Instruct-Turbo ($0.88/$0.88, from LiteLLM together_ai source). Cheaper on DeepInfra ($0.13/$0.39); Vertex $0.72/$0.72."
+  },
+  "command-a": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://cohere.com/pricing",
+    "as_of": "2026-06",
+    "_note": "command-a-03-2025."
+  },
+  "command-r-plus": {
+    "input_per_mtok": 2.5,
+    "output_per_mtok": 10,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://cohere.com/pricing",
+    "as_of": "2026-06",
+    "_note": "command-r-plus-08-2024 (legacy)."
+  },
+  "amazon-nova-pro": {
+    "input_per_mtok": 0.8,
+    "output_per_mtok": 3.2,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://aws.amazon.com/bedrock/pricing/",
+    "as_of": "2026-06",
+    "_note": "AWS Bedrock on-demand ($0.0008/$0.0032 per 1K tokens)."
+  },
+  "amazon-nova-lite": {
+    "input_per_mtok": 0.06,
+    "output_per_mtok": 0.24,
+    "cache_read_per_mtok": null,
+    "cache_write_per_mtok": null,
+    "source": "https://aws.amazon.com/bedrock/pricing/",
+    "as_of": "2026-06",
+    "_note": "AWS Bedrock on-demand ($0.00006/$0.00024 per 1K tokens)."
+  }
+}

src/session/extract.ts

@@ -5,6 +5,11 @@
  * All 13 event categories as specified in PRD Section 3.
  */
 
+import {
+  lookupPrice as catalogLookupPrice,
+  computeCostUsd as catalogComputeCostUsd,
+} from "../pricing/catalog.js";
+
 // ── Public interfaces ──────────────────────────────────────────────────────
 
 export interface SessionEvent {
@@ -1339,62 +1344,92 @@ function extractFileReadMetadata(input: HookInput): SessionEvent[] {
 }
 
 /**
- * Per-model USD price table — Anthropic public list pricing, $/MTok.
- * Verified against platform.claude.com/docs/en/about-claude/pricing,
- * cloudzero.com, finout.io 2026-06 (cache: 5-min cache_write = 1.25× input,
- * cache_read = 0.10× input). Fast-mode variants (e.g. opus-4-8-fast at
- * $10/$50) are intentionally NOT mapped — they ship as separate model
- * ids and would dilute the standard-tier dashboards if blended here.
+ * Per-model USD pricing now lives in the curated multi-vendor catalog
+ * (src/pricing/catalog.ts), which prices each model from ITS OWN row across
+ * Anthropic / OpenAI / Google / Chinese / other vendors. This kills the old
+ * bug where the hardcoded Anthropic-only table here billed every non-Claude
+ * model at Claude-Sonnet's `default` rate. Unknown ids now resolve to a null
+ * cost (one console.warn) instead of a silently wrong Claude rate.
  *
- * NOTE: 16-oss-verify-gap-prd Gap #1 quoted Opus at $15/$75 — that is
- * the prior Opus 4 (non-4.7) rate. Opus 4.7 and 4.8 ship at $5/$25.
+ * resolveModelId picks the first non-empty model id from the hook candidates;
+ * date-suffixed ids (e.g. claude-haiku-4-5-20251001) are reduced to a catalog
+ * hit by progressively dropping trailing `-segment` suffixes (NO regex).
  */
-const MODEL_PRICING_USD_PER_MTOK: Record<string, {
-  input: number;
-  output: number;
-  cache_write: number;
-  cache_read: number;
-}> = {
-  "claude-opus-4-8":   { input: 5.00, output: 25.00, cache_write: 6.25, cache_read: 0.50 },
-  "claude-opus-4-7":   { input: 5.00, output: 25.00, cache_write: 6.25, cache_read: 0.50 },
-  "claude-sonnet-4-6": { input: 3.00, output: 15.00, cache_write: 3.75, cache_read: 0.30 },
-  "claude-haiku-4-5":  { input: 1.00, output:  5.00, cache_write: 1.25, cache_read: 0.10 },
-  default:             { input: 3.00, output: 15.00, cache_write: 3.75, cache_read: 0.30 },
-};
-
-function resolveModelKey(input: HookInput, parsedResp: Record<string, unknown>): string {
+function resolveModelId(input: HookInput, parsedResp: Record<string, unknown>): string {
   const candidates: unknown[] = [
     input.tool_input?.model,
     (input as unknown as Record<string, unknown>).model,
     parsedResp.model,
   ];
-  const keys = Object.keys(MODEL_PRICING_USD_PER_MTOK).filter((k) => k !== "default");
   for (const c of candidates) {
-    if (typeof c !== "string" || c.length === 0) continue;
-    if (c in MODEL_PRICING_USD_PER_MTOK) return c;
-    // Prefix match for date-suffixed model ids
-    // (e.g. claude-haiku-4-5-20251001 → claude-haiku-4-5)
-    for (const key of keys) {
-      if (c.startsWith(key)) return key;
-    }
+    if (typeof c === "string" && c.length > 0) return c;
   }
-  return "default";
+  return "";
 }
 
-function computeCostUsd(
-  modelKey: string,
+/**
+ * Drop one trailing `-<segment>` from a model id, char-algorithmically (no
+ * regex): walks back to the last '-' and returns the head, or null when there
+ * is no usable separator. Lets a date-suffixed id fall back to its base id
+ * (claude-haiku-4-5-20251001 → claude-haiku-4-5 → … ) one segment at a time.
+ */
+function dropTrailingSegment(id: string): string | null {
+  for (let i = id.length - 1; i > 0; i--) {
+    if (id.charCodeAt(i) === 45 /* '-' */) return id.slice(0, i);
+  }
+  return null;
+}
+
+/**
+ * Resolve a model id to one the catalog can price: try the raw id, then
+ * progressively trim trailing `-segment` suffixes so a date-suffixed id still
+ * prices off its base model. Probes with lookupPrice (no warn) and returns the
+ * first id that hits, or "" on a full miss — so cost compute warns at most once.
+ */
+function resolveCatalogId(modelId: string): string {
+  let candidate: string | null = modelId;
+  while (candidate && candidate.length > 0) {
+    if (catalogLookupPrice(candidate) !== null) return candidate;
+    candidate = dropTrailingSegment(candidate);
+  }
+  return "";
+}
+
+/**
+ * Cost for a turn via the catalog. Returns null on a price miss (catalog emits
+ * one console.warn of the unmatched id) or when all token buckets are zero.
+ */
+function computeTurnCostUsd(
+  modelId: string,
   inputTokens: number,
   outputTokens: number,
   cacheCreationTokens: number,
   cacheReadTokens: number,
-): number {
-  const price = MODEL_PRICING_USD_PER_MTOK[modelKey] ?? MODEL_PRICING_USD_PER_MTOK.default;
-  const totalMicroDollars =
-    inputTokens * price.input +
-    outputTokens * price.output +
-    cacheCreationTokens * price.cache_write +
-    cacheReadTokens * price.cache_read;
-  return totalMicroDollars / 1_000_000;
+): number | null {
+  const resolved = resolveCatalogId(modelId);
+  // Feed the resolved id when found; otherwise pass the raw id so the catalog's
+  // single miss-warning carries the id the operator actually saw.
+  return catalogComputeCostUsd(resolved || modelId, {
+    input_tokens: inputTokens,
+    output_tokens: outputTokens,
+    cache_creation_tokens: cacheCreationTokens,
+    cache_read_tokens: cacheReadTokens,
+  });
+}
+
+/**
+ * Format a cost to a compact `cost_usd` string, char-algorithmically (no
+ * regex). Renders 6 decimals, drops trailing zeros, and keeps a single `.0`
+ * when the fraction trims to empty (e.g. 0 → "0.0"), matching the prior
+ * `.toFixed(6).replace(...)` output exactly.
+ */
+function formatCostUsd(cost: number): string {
+  let s = cost.toFixed(6);
+  let end = s.length;
+  while (end > 0 && s.charCodeAt(end - 1) === 48 /* '0' */) end--;
+  s = s.slice(0, end);
+  if (s.length > 0 && s.charCodeAt(s.length - 1) === 46 /* '.' */) s += "0";
+  return s;
 }
 
 /**
@@ -1454,9 +1489,11 @@ function extractAgentUsage(input: HookInput): SessionEvent[] {
     : 0;
   const anyTokens = inputTokens > 0 || outputTokens > 0 || cacheCreate > 0 || cacheRead > 0;
   if (anyTokens) {
-    const modelKey = resolveModelKey(input, out);
-    const cost = computeCostUsd(modelKey, inputTokens, outputTokens, cacheCreate, cacheRead);
-    parts.push(`cost_usd:${cost.toFixed(6).replace(/0+$/, "").replace(/\.$/, ".0")}`);
+    const modelId = resolveModelId(input, out);
+    // null ⇒ unmatched model id (catalog warned once) — skip the cost token
+    // rather than blend a wrong Claude rate (the old non-Claude bug).
+    const cost = computeTurnCostUsd(modelId, inputTokens, outputTokens, cacheCreate, cacheRead);
+    if (cost !== null) parts.push(`cost_usd:${formatCostUsd(cost)}`);
   }
 
   return [{

tests/pricing/catalog.test.ts

@@ -0,0 +1,177 @@
+/**
+ * Pricing catalog — curated multi-vendor price lookup + cost compute.
+ *
+ * Replaces the Anthropic-only hardcoded table that lived in
+ * src/session/extract.ts (≈1356-1396). That table priced EVERY model at
+ * a Claude rate (non-Claude models silently inherited Sonnet's default),
+ * which over-/under-charged every OpenAI / Gemini / Qwen / DeepSeek turn.
+ *
+ * The catalog merges the 5 curated vendor JSONs (src/pricing/sources/*.json)
+ * into one per-Mtok price map and prices each model from ITS OWN row.
+ * Behaviours under test:
+ *   (a) curated lookup hits across all five vendors
+ *   (b) provider/ prefix strip (anthropic/claude-opus-4-8 → claude-opus-4-8)
+ *   (c) non-Claude model gets ITS price, not Claude's (the bug)
+ *   (d) unknown id → null + one console.warn of the unmatched id
+ *   (e) cache_read vs cache_creation weighting (distinct prices)
+ *   (f) all-zero / absent tokens → null
+ *   (g) native-cost passthrough wins over computed
+ */
+
+import { describe, test, expect, vi } from "vitest";
+import {
+  lookupPrice,
+  computeCostUsd,
+  nativeOrComputed,
+} from "../../src/pricing/catalog.js";
+
+describe("pricing/catalog — lookupPrice", () => {
+  // (a) curated lookup hits
+  test("(a) Anthropic curated hit returns per-Mtok price", () => {
+    const p = lookupPrice("claude-opus-4-8");
+    expect(p).not.toBeNull();
+    expect(p!.input_per_mtok).toBe(5);
+    expect(p!.output_per_mtok).toBe(25);
+    expect(p!.cache_write_per_mtok).toBe(6.25);
+    expect(p!.cache_read_per_mtok).toBe(0.5);
+  });
+
+  test("(a) OpenAI curated hit", () => {
+    expect(lookupPrice("gpt-5")!.input_per_mtok).toBe(1.25);
+  });
+
+  test("(a) Google curated hit", () => {
+    expect(lookupPrice("gemini-2.5-flash")!.output_per_mtok).toBe(2.5);
+  });
+
+  test("(a) Chinese vendor curated hit", () => {
+    expect(lookupPrice("deepseek-v3")).not.toBeNull();
+  });
+
+  test("(a) Other vendor curated hit", () => {
+    expect(lookupPrice("grok-4")).not.toBeNull();
+  });
+
+  // (b) provider/ prefix strip + normalization
+  test("(b) provider/ prefix is stripped", () => {
+    const stripped = lookupPrice("anthropic/claude-opus-4-8");
+    const bare = lookupPrice("claude-opus-4-8");
+    expect(stripped).toEqual(bare);
+  });
+
+  test("(b) normalization trims and lowercases", () => {
+    expect(lookupPrice("  GPT-5  ")).toEqual(lookupPrice("gpt-5"));
+  });
+
+  test("(b) openrouter-style provider/model strips first segment only", () => {
+    // openai/gpt-5 → gpt-5
+    expect(lookupPrice("openai/gpt-5")).toEqual(lookupPrice("gpt-5"));
+  });
+
+  // (d) unknown id → null
+  test("(d) unknown model id returns null", () => {
+    expect(lookupPrice("totally-made-up-model-xyz")).toBeNull();
+  });
+
+  // null-priced curated entry → treated as no price
+  test("null-priced curated entry (gemini-2.0-pro) is treated as no price", () => {
+    expect(lookupPrice("gemini-2.0-pro")).toBeNull();
+  });
+});
+
+describe("pricing/catalog — computeCostUsd", () => {
+  // (c) THE BUG: non-Claude model must use its own price, not Claude's.
+  test("(c) gpt-5 priced from its own row, NOT Claude default", () => {
+    const tokens = { input_tokens: 1000, output_tokens: 500 };
+    const gpt5 = computeCostUsd("gpt-5", tokens);
+    // gpt-5: 1000*1.25 + 500*10 = 1250 + 5000 = 6250 / 1e6 = 0.00625
+    expect(gpt5).toBeCloseTo(0.00625, 8);
+    // Claude-default (old buggy path) would have produced Sonnet 0.0105.
+    const sonnet = computeCostUsd("claude-sonnet-4-6", tokens);
+    expect(gpt5).not.toBeCloseTo(sonnet!, 8);
+  });
+
+  test("(c) gemini-2.5-flash priced from its own row", () => {
+    // 1000*0.3 + 500*2.5 = 300 + 1250 = 1550 / 1e6 = 0.00155
+    expect(computeCostUsd("gemini-2.5-flash", {
+      input_tokens: 1000,
+      output_tokens: 500,
+    })).toBeCloseTo(0.00155, 8);
+  });
+
+  // opus-4-8 regression — must match the old hardcoded table exactly.
+  test("opus-4-8 regression: 1000 in + 500 out = 0.0175", () => {
+    expect(computeCostUsd("claude-opus-4-8", {
+      input_tokens: 1000,
+      output_tokens: 500,
+    })).toBeCloseTo(0.0175, 8);
+  });
+
+  // (e) cache_read vs cache_creation weighting
+  test("(e) cache_creation uses cache_write price, cache_read uses cache_read price", () => {
+    // sonnet: in 3, out 15, cache_write 3.75, cache_read 0.30
+    // 1000*3 + 500*15 + 1000*3.75 + 1500*0.30 = 3000+7500+3750+450 = 14700
+    expect(computeCostUsd("claude-sonnet-4-6", {
+      input_tokens: 1000,
+      output_tokens: 500,
+      cache_creation_tokens: 1000,
+      cache_read_tokens: 1500,
+    })).toBeCloseTo(0.0147, 8);
+  });
+
+  test("(e) null cache price falls back to input price for that bucket", () => {
+    // gpt-5 has cache_write_per_mtok=null → cache_creation billed at input (1.25).
+    // 2000 cache_creation tokens only: 2000*1.25 = 2500 / 1e6 = 0.0025
+    expect(computeCostUsd("gpt-5", {
+      cache_creation_tokens: 2000,
+    })).toBeCloseTo(0.0025, 8);
+  });
+
+  // (f) all-zero / absent tokens → null
+  test("(f) all-zero tokens returns null", () => {
+    expect(computeCostUsd("claude-sonnet-4-6", {
+      input_tokens: 0,
+      output_tokens: 0,
+      cache_read_tokens: 0,
+      cache_creation_tokens: 0,
+    })).toBeNull();
+  });
+
+  test("(f) absent tokens returns null", () => {
+    expect(computeCostUsd("claude-sonnet-4-6", {})).toBeNull();
+  });
+
+  // (d) unknown model → null + single console.warn of the unmatched id
+  test("(d) unknown model returns null and warns once with the id", () => {
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => {});
+    try {
+      const cost = computeCostUsd("unknown-vendor/mystery-9000", {
+        input_tokens: 1000,
+      });
+      expect(cost).toBeNull();
+      expect(warn).toHaveBeenCalledTimes(1);
+      expect(String(warn.mock.calls[0]?.[0])).toContain("mystery-9000");
+    } finally {
+      warn.mockRestore();
+    }
+  });
+});
+
+describe("pricing/catalog — nativeOrComputed", () => {
+  // (g) native-cost passthrough
+  test("(g) provider native cost wins over computed", () => {
+    const native = nativeOrComputed("gpt-5", { input_tokens: 1000 }, 0.42);
+    expect(native).toBe(0.42);
+  });
+
+  test("(g) falls back to computed when native is null/undefined", () => {
+    const computed = computeCostUsd("gpt-5", { input_tokens: 1000 });
+    expect(nativeOrComputed("gpt-5", { input_tokens: 1000 }, null)).toBe(computed);
+    expect(nativeOrComputed("gpt-5", { input_tokens: 1000 })).toBe(computed);
+  });
+
+  test("(g) native cost of 0 is a real value, not a miss", () => {
+    // A provider that genuinely charged $0 (free tier) must pass through.
+    expect(nativeOrComputed("gpt-5", { input_tokens: 1000 }, 0)).toBe(0);
+  });
+});

tests/session/extract-agent-usage.test.ts

@@ -19,7 +19,7 @@
  * forward-compatible Zod envelope accepts them today (no migration).
  */
 
-import { describe, test, expect } from "vitest";
+import { describe, test, expect, vi } from "vitest";
 import { extractEvents } from "../../src/session/extract.js";
 
 function agentUsageOf(toolResponse: unknown, toolName: string = "Task") {
@@ -211,7 +211,12 @@ describe("extractAgentUsage — Issue #4 AgentOutput.usage capture", () => {
     expect(events[0].data).toMatch(/cost_usd:0\.0147/);
   });
 
-  test("cost_usd: unknown model falls back to default Sonnet pricing", () => {
+  test("cost_usd: unknown model → no cost emitted (was: wrong Claude fallback)", () => {
+    // Catalog rewire — an unmatched model id no longer inherits Claude-Sonnet's
+    // rate (the old `default` bug). It resolves to null cost, so the event is
+    // still emitted (token counts) but carries NO cost_usd token. The catalog
+    // also console.warns the unmatched id once; silence it here.
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => {});
     const events = extractEvents({
       tool_name: "Task",
       tool_input: { model: "claude-future-model-99" },
@@ -220,7 +225,9 @@ describe("extractAgentUsage — Issue #4 AgentOutput.usage capture", () => {
         usage: { input_tokens: 1000, output_tokens: 500 },
       }),
     }).filter((e) => e.type === "agent_usage");
-    expect(events[0].data).toMatch(/cost_usd:0\.0105/);
+    warn.mockRestore();
+    expect(events.length).toBe(1);
+    expect(events[0].data).not.toMatch(/cost_usd:/);
   });
 
   test("cost_usd: Opus 4.8 priced at same standard rate as Opus 4.7", () => {
@@ -255,7 +262,10 @@ describe("extractAgentUsage — Issue #4 AgentOutput.usage capture", () => {
     expect(events[0].data).toMatch(/cost_usd:0\.0035/);
   });
 
-  test("cost_usd: no model + token counts → still computes with default pricing", () => {
+  test("cost_usd: no model id → no cost emitted (cannot price without a model)", () => {
+    // Catalog rewire — without a model id there is no row to price from, so no
+    // cost_usd is blended. The empty id is not warned (it is not a real miss).
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => {});
     const events = extractEvents({
       tool_name: "Task",
       tool_input: {},
@@ -263,7 +273,9 @@ describe("extractAgentUsage — Issue #4 AgentOutput.usage capture", () => {
         usage: { input_tokens: 1000, output_tokens: 500 },
       }),
     }).filter((e) => e.type === "agent_usage");
-    expect(events[0].data).toMatch(/cost_usd:/);
+    warn.mockRestore();
+    expect(events.length).toBe(1);
+    expect(events[0].data).not.toMatch(/cost_usd:/);
   });
 
   test("cost_usd: zero tokens → cost_usd:0 not emitted (skip)", () => {

tools/pricing/litellm-NOTES.md

@@ -0,0 +1,80 @@
+# LiteLLM Catalog — Adapter Notes
+
+Vendored from BerriAI/litellm as context-mode's comprehensive pricing base, so
+unknown/unseen models still resolve a price instead of failing.
+
+- **Source:** `https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json`
+- **Local file:** `litellm-catalog.json`
+- **Size:** ~1.53 MB (1,570,159 bytes)
+- **Top-level keys:** 2,901 (2,900 model entries + 1 `sample_spec` template — skip `sample_spec`)
+- **Entries with numeric `input_cost_per_token`:** 2,432
+- **Distinct `litellm_provider` values:** 120 (top: fireworks_ai, bedrock, openai, azure, gemini, mistral, openrouter …)
+
+## Shape
+
+The file is a flat JSON object. **Each key is a model id**; each value is a metadata object.
+There is no wrapper array. One special key, `sample_spec`, is a documentation template
+(not a real model) and must be excluded.
+
+```jsonc
+{
+  "sample_spec": { /* template — ignore */ },
+  "gpt-4o": { "input_cost_per_token": 0.0000025, ... },
+  "claude-sonnet-4-20250514": { ... }
+}
+```
+
+## Cost fields — verified present in the fetched JSON
+
+> All `*_cost_per_token` values are **PER TOKEN** (USD), not per-million.
+> To convert to per-Mtok (context-mode's internal unit): **multiply by `1e6` (×1,000,000).**
+> e.g. `gpt-4o` `input_cost_per_token = 0.0000025` → `0.0000025 × 1e6 = $2.50 / Mtok`.
+
+Core fields (use these; coverage count in parentheses):
+
+| Field | Meaning | Count |
+|-------|---------|-------|
+| `input_cost_per_token` | prompt cost per token | 2,433 |
+| `output_cost_per_token` | completion cost per token | 2,430 |
+| `cache_read_input_token_cost` | cached-prompt read cost per token | 645 |
+| `cache_creation_input_token_cost` | cache-write/creation cost per token | 210 |
+
+Real example (`claude-sonnet-4-20250514`, provider `anthropic`):
+`input_cost_per_token: 0.000003` (= $3.00/Mtok), `output_cost_per_token: 0.000015` (= $15.00/Mtok),
+`cache_read_input_token_cost: 3e-7` (= $0.30/Mtok), `cache_creation_input_token_cost: 0.00000375` (= $3.75/Mtok).
+
+### Schema surprises / gotchas
+
+- **Tiered & variant cost fields exist** — e.g. `input_cost_per_token_above_200k_tokens`,
+  `cache_read_input_token_cost_above_200k_tokens`, `_priority`, `_batches`, `_flex`,
+  `_above_1hr`, `_above_272k_tokens`. Treat these as optional overrides; fall back to the
+  base `input_cost_per_token` / `output_cost_per_token`.
+- **Non-token cost units also appear** and are NOT per-token: `input_cost_per_second`,
+  `input_cost_per_character`, `input_cost_per_image`, `input_cost_per_pixel`,
+  `output_cost_per_second`, `output_cost_per_reasoning_token`,
+  `search_context_cost_per_query` (the last can be an object keyed by search-context size).
+  Do not blindly ×1e6 these — only the `*_per_token` family is per-token.
+- **2,366 of 2,900 keys contain `/`** — namespaced ids like `bedrock/...`,
+  `1024-x-1024/dall-e-2`, image/resolution-prefixed entries. Match on the full key.
+- Metadata fields used for context limits: `max_tokens`, `max_input_tokens`,
+  `max_output_tokens` (and `mode` distinguishes `chat`, `embedding`, image, etc.).
+- Some entries carry `deprecation_date` (81 entries).
+
+## Lookup strategy (for the adapter)
+
+Given an incoming `model_id`:
+
+1. **Exact match** — `catalog[model_id]`. Fastest; covers the common case.
+2. **Provider-stripped / namespaced fallback** — many ids are `provider/model`.
+   Try stripping or adding a known provider prefix:
+   - if `model_id` has no `/`, try `catalog[provider + "/" + model_id]`;
+   - if `model_id` is `provider/model`, also try the bare `model` segment.
+3. **Provider + model match** — scan entries whose `litellm_provider` matches the resolved
+   provider and whose key endsWith the model segment.
+4. **Skip `sample_spec`** in every scan.
+5. From the matched entry, read `input_cost_per_token` / `output_cost_per_token`
+   (+ optional `cache_read_input_token_cost`, `cache_creation_input_token_cost`),
+   then **× 1e6** to get per-Mtok rates.
+6. If nothing matches or `input_cost_per_token` is absent (some entries price only by
+   second/character/image), the model has no usable per-token price — fall through to
+   context-mode's own default.