@broberg/ai-sdk 0.30.1 → 0.32.0

package/dist/{chunk-LKVCPMVI.js → chunk-DEIY7O3T.js}

@@ -61,18 +61,24 @@ var PRICING = {
   // Verify against a real key when it lands.
   "deepseek:deepseek-chat": { inputPer1M: 0.14, outputPer1M: 0.28, version: "2026-06-30-deepseek-direct" },
   "deepseek:deepseek-reasoner": { inputPer1M: 0.14, outputPer1M: 0.28, version: "2026-06-30-deepseek-direct" },
+  // Cached input tokens cost 10% of the input rate — $0.03 vs $0.30 (2.5-flash) and
+  // $0.01 vs $0.10 (2.5-flash-lite), read from ai.google.dev/gemini-api/docs/pricing
+  // on 2026-08-27 rather than recalled. NB the storage fee on that page ($1/1M
+  // tokens/hour) applies to EXPLICIT context caching, where you create a CachedContent
+  // object with a TTL. We use IMPLICIT caching, which has no storage charge — so this
+  // table is not silently under-billing.
   // Google Gemini (direct). Provider key is "gemini" — matches the adapter's
   // usage.provider + the override.provider callers pass. (Image-gen models are
   // priced per-image in the adapter, not here.)
-  "gemini:gemini-2.5-flash": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },
+  "gemini:gemini-2.5-flash": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 2.5, version: "2026-08-27-ai.google.dev" },
   // flash-lite is the default `video` tier (F019) — cheap native video understanding.
-  "gemini:gemini-2.5-flash-lite": { inputPer1M: 0.1, outputPer1M: 0.4, version: "2026-06-04-or-xref" },
+  "gemini:gemini-2.5-flash-lite": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: "2026-08-27-ai.google.dev" },
   // Vertex AI (F038) — the EU-resident route to the SAME Gemini models, so Google's
   // published Gemini token prices apply. Listed separately because cost lookups key on
   // `provider:model`: without these rows an EU vision/video call would silently log
   // $0, which is worse than no tracking (a confident wrong number).
-  "vertex:gemini-2.5-flash": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },
-  "vertex:gemini-2.5-flash-lite": { inputPer1M: 0.1, outputPer1M: 0.4, version: "2026-06-04-or-xref" },
+  "vertex:gemini-2.5-flash": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 2.5, version: "2026-08-27-ai.google.dev" },
+  "vertex:gemini-2.5-flash-lite": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: "2026-08-27-ai.google.dev" },
   // Cached prompt tokens bill at 10% of the input rate (F039, measured 2026-08-27:
   // an 8,810-token prefix reported 8,784 cached on the second call WITH a
   // prompt_cache_key, and 0 without one at every size up to 57k).
@@ -114,4 +120,4 @@ export {
   PRICING,
   getPrice
 };
-//# sourceMappingURL=chunk-LKVCPMVI.js.map
+//# sourceMappingURL=chunk-DEIY7O3T.js.map

package/dist/chunk-DEIY7O3T.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/cost/pricing.ts"],"sourcesContent":["// Versioned per-(provider, model) pricing. F3.6 populates the table + adds tests\n// + MiniMax coverage. F3.1 ships the type + lookup with an empty table, so\n// computeCost returns 0 for every model until F3.6 lands (calls still complete).\nexport interface PricingEntry {\n  /** USD per 1M input tokens. */\n  inputPer1M: number;\n  /** USD per 1M output tokens. */\n  outputPer1M: number;\n  /** USD per 1M cache-read tokens (falls back to input rate if unset). */\n  cacheReadPer1M?: number;\n  /** USD per 1M cache-write/creation tokens (falls back to input rate if unset). */\n  cacheWritePer1M?: number;\n  /** Pricing snapshot version (date or tag) so stale entries are detectable. */\n  version: string;\n}\n\n// USD per 1M tokens. Anthropic cache multipliers follow the standard model:\n// cache-read ≈ 0.1× input, cache-write ≈ 1.25× input. Verified against the\n// pricing tables in cms (packages/cms-ai/src/providers) + trail (model-lab).\n// MiniMax M2.7 is an estimate pending confirmation against OpenRouter's live\n// price page — flagged in its version string.\nconst V = \"2026-06-02\";\n// Mistral prices come straight from mistral.ai/pricing (per Christian's CD report).\nconst MS = \"2026-06-04-mistral.ai\";\n\n/** Keyed `${provider}:${model}`. Exported so the catalogue-research job (F014)\n *  can enumerate every priced entry and diff it against the live provider lists. */\nexport const PRICING: Record<string, PricingEntry> = {\n  // Anthropic (direct API). DEFAULT_TIER_MAP: fast/cheap=haiku, smart/vision=sonnet, powerful=opus.\n  \"anthropic:claude-haiku-4-5\": {\n    inputPer1M: 0.8,\n    outputPer1M: 4.0,\n    cacheReadPer1M: 0.08,\n    cacheWritePer1M: 1.0,\n    version: V,\n  },\n  \"anthropic:claude-sonnet-4-6\": {\n    inputPer1M: 3.0,\n    outputPer1M: 15.0,\n    cacheReadPer1M: 0.3,\n    cacheWritePer1M: 3.75,\n    version: V,\n  },\n  \"anthropic:claude-opus-4-8\": {\n    inputPer1M: 15.0,\n    outputPer1M: 75.0,\n    cacheReadPer1M: 1.5,\n    cacheWritePer1M: 18.75,\n    version: V,\n  },\n\n  // OpenAI. embedding default tier = text-embedding-3-small (no output tokens).\n  \"openai:text-embedding-3-small\": { inputPer1M: 0.02, outputPer1M: 0, version: V },\n  \"openai:text-embedding-3-large\": { inputPer1M: 0.13, outputPer1M: 0, version: V },\n  \"openai:gpt-4o\": { inputPer1M: 2.5, outputPer1M: 10.0, version: V },\n  \"openai:gpt-4o-mini\": { inputPer1M: 0.15, outputPer1M: 0.6, version: V },\n  // Whisper is priced per minute, not per token — not representable here; transcribe\n  // (F5.6) computes its own cost. Listed as 0 so token-based compute never charges it.\n  \"openai:whisper-1\": { inputPer1M: 0, outputPer1M: 0, version: V },\n\n  // OpenRouter (meta-router — model slugs include the upstream vendor). Slugs use\n  // dots (claude-sonnet-4.6) to match OpenRouter's live ids; the dashed forms\n  // never matched a real call. Caught by the F014 catalogue research.\n  \"openrouter:anthropic/claude-sonnet-4.6\": { inputPer1M: 3.0, outputPer1M: 15.0, version: V },\n  // OpenRouter ground-truth $1/$5 — a markup over Anthropic-direct's $0.8/$4\n  // (the `anthropic:` entry above). Was masked while the slug used dashes.\n  \"openrouter:anthropic/claude-haiku-4.5\": { inputPer1M: 1.0, outputPer1M: 5.0, version: \"2026-06-04\" },\n  \"openrouter:google/gemini-2.5-flash\": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },\n  // Ground-truth from OpenRouter /api/v1/models (was a 0.3 estimate; now 0.279).\n  \"openrouter:minimax/minimax-m2.7\": {\n    inputPer1M: 0.279,\n    outputPer1M: 1.2,\n    version: \"2026-06-04\",\n  },\n  // DeepSeek V4 (CN-hosted — NOT GDPR-safe; non-personal-data workloads only).\n  // On 2026-05-22 DeepSeek made the \"75% off\" promo the permanent official price.\n  // V4-Pro $0.435/$0.87 is ~34x cheaper than GPT-5.5 on output; flash is cheaper\n  // still. Numbers match OpenRouter /api/v1/models 1:1 (no router markup). A strong\n  // cheap route for fleet background work once `claude -p` is API-billed (15 Jun).\n  \"openrouter:deepseek/deepseek-v4-pro\": { inputPer1M: 0.435, outputPer1M: 0.87, version: \"2026-05-22-deepseek-official\" },\n  \"openrouter:deepseek/deepseek-v4-flash\": { inputPer1M: 0.0983, outputPer1M: 0.1966, version: \"2026-05-22-deepseek-official\" },\n  // DeepSeek DIRECT API (provider \"deepseek\", F030 non-PII secondary). Rates from\n  // api-docs.deepseek.com 2026-06-30 ($0.14/$0.28 per 1M; both map to deepseek-v4-flash).\n  // `deepseek-chat` (non-thinking) + `deepseek-reasoner` (thinking) DEPRECATE 2026-07-24.\n  // (The bare `deepseek-v4-flash` basename is already priced via the openrouter entry\n  // above — kept distinct here to avoid a basename collision in the F027 pricing-API.)\n  // Verify against a real key when it lands.\n  \"deepseek:deepseek-chat\": { inputPer1M: 0.14, outputPer1M: 0.28, version: \"2026-06-30-deepseek-direct\" },\n  \"deepseek:deepseek-reasoner\": { inputPer1M: 0.14, outputPer1M: 0.28, version: \"2026-06-30-deepseek-direct\" },\n\n  // Cached input tokens cost 10% of the input rate — $0.03 vs $0.30 (2.5-flash) and\n  // $0.01 vs $0.10 (2.5-flash-lite), read from ai.google.dev/gemini-api/docs/pricing\n  // on 2026-08-27 rather than recalled. NB the storage fee on that page ($1/1M\n  // tokens/hour) applies to EXPLICIT context caching, where you create a CachedContent\n  // object with a TTL. We use IMPLICIT caching, which has no storage charge — so this\n  // table is not silently under-billing.\n  // Google Gemini (direct). Provider key is \"gemini\" — matches the adapter's\n  // usage.provider + the override.provider callers pass. (Image-gen models are\n  // priced per-image in the adapter, not here.)\n  \"gemini:gemini-2.5-flash\": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 2.5, version: \"2026-08-27-ai.google.dev\" },\n  // flash-lite is the default `video` tier (F019) — cheap native video understanding.\n  \"gemini:gemini-2.5-flash-lite\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: \"2026-08-27-ai.google.dev\" },\n\n  // Vertex AI (F038) — the EU-resident route to the SAME Gemini models, so Google's\n  // published Gemini token prices apply. Listed separately because cost lookups key on\n  // `provider:model`: without these rows an EU vision/video call would silently log\n  // $0, which is worse than no tracking (a confident wrong number).\n  \"vertex:gemini-2.5-flash\": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 2.5, version: \"2026-08-27-ai.google.dev\" },\n  \"vertex:gemini-2.5-flash-lite\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: \"2026-08-27-ai.google.dev\" },\n\n  // Cached prompt tokens bill at 10% of the input rate (F039, measured 2026-08-27:\n  // an 8,810-token prefix reported 8,784 cached on the second call WITH a\n  // prompt_cache_key, and 0 without one at every size up to 57k).\n  // Mistral (direct, La Plateforme). Official prices from mistral.ai/pricing\n  // (2026-06-04, per Christian's CD report). EU/Paris-hosted — the designated\n  // GDPR-safe provider for client/personal-data workloads (see F015). NB:\n  // medium-3.5 is the premium \"Vibe\" coding tier ($1.5/$7.5); Large 3 ($0.5/$1.5)\n  // is the cheaper frontier general-purpose model despite the higher number.\n  \"mistral:mistral-large-latest\": { inputPer1M: 0.5, cacheReadPer1M: 0.05, outputPer1M: 1.5, version: MS },\n  \"mistral:mistral-large-2512\": { inputPer1M: 0.5, cacheReadPer1M: 0.05, outputPer1M: 1.5, version: MS },\n  \"mistral:mistral-medium-latest\": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: MS },\n  \"mistral:mistral-medium-3.5\": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: MS },\n  \"mistral:mistral-medium-3\": { inputPer1M: 0.4, cacheReadPer1M: 0.04, outputPer1M: 2.0, version: \"2026-06-04-or-xref\" },\n  \"mistral:mistral-small-latest\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.3, version: MS },\n  \"mistral:mistral-small-2603\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.3, version: MS },\n  \"mistral:ministral-3b-latest\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.1, version: MS },\n  \"mistral:ministral-8b-latest\": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.15, version: MS },\n  \"mistral:ministral-14b-latest\": { inputPer1M: 0.2, cacheReadPer1M: 0.02, outputPer1M: 0.2, version: MS },\n  \"mistral:magistral-medium-latest\": { inputPer1M: 2.0, cacheReadPer1M: 0.2, outputPer1M: 5.0, version: MS },\n  \"mistral:magistral-small-latest\": { inputPer1M: 0.5, cacheReadPer1M: 0.05, outputPer1M: 1.5, version: MS },\n  \"mistral:devstral-latest\": { inputPer1M: 0.4, cacheReadPer1M: 0.04, outputPer1M: 2.0, version: MS },\n  \"mistral:codestral-latest\": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 0.9, version: MS },\n  \"mistral:open-mistral-nemo\": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.15, version: MS },\n  // Moderation (F016.4) — per input token; output 0. (OCR is per-page in the adapter.)\n  \"mistral:mistral-moderation-latest\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0, version: MS },\n  // Embeddings (F016.5) — per input token.\n  \"mistral:mistral-embed\": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0, version: MS },\n  \"mistral:codestral-embed\": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0, version: MS },\n};\n\nexport function getPrice(provider: string, model: string): PricingEntry | undefined {\n  const exact = PRICING[`${provider}:${model}`];\n  if (exact) return exact;\n  // Providers ship dated model snapshots, e.g. \"claude-haiku-4-5-20251001\".\n  // Strip a trailing -YYYYMMDD and retry the base lookup so a dated variant\n  // prices the same as its base model instead of falling through to 0 — a real\n  // paid call must never be logged as $0 (F012). Covers openrouter slugs too.\n  const base = model.replace(/-\\d{8}$/, \"\");\n  if (base !== model) return PRICING[`${provider}:${base}`];\n  return undefined;\n}\n"],"mappings":";AAqBA,IAAM,IAAI;AAEV,IAAM,KAAK;AAIJ,IAAM,UAAwC;AAAA;AAAA,EAEnD,8BAA8B;AAAA,IAC5B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA,EACA,+BAA+B;AAAA,IAC7B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA,EACA,6BAA6B;AAAA,IAC3B,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,gBAAgB;AAAA,IAChB,iBAAiB;AAAA,IACjB,SAAS;AAAA,EACX;AAAA;AAAA,EAGA,iCAAiC,EAAE,YAAY,MAAM,aAAa,GAAG,SAAS,EAAE;AAAA,EAChF,iCAAiC,EAAE,YAAY,MAAM,aAAa,GAAG,SAAS,EAAE;AAAA,EAChF,iBAAiB,EAAE,YAAY,KAAK,aAAa,IAAM,SAAS,EAAE;AAAA,EAClE,sBAAsB,EAAE,YAAY,MAAM,aAAa,KAAK,SAAS,EAAE;AAAA;AAAA;AAAA,EAGvE,oBAAoB,EAAE,YAAY,GAAG,aAAa,GAAG,SAAS,EAAE;AAAA;AAAA;AAAA;AAAA,EAKhE,0CAA0C,EAAE,YAAY,GAAK,aAAa,IAAM,SAAS,EAAE;AAAA;AAAA;AAAA,EAG3F,yCAAyC,EAAE,YAAY,GAAK,aAAa,GAAK,SAAS,aAAa;AAAA,EACpG,sCAAsC,EAAE,YAAY,KAAK,aAAa,KAAK,SAAS,EAAE;AAAA;AAAA,EAEtF,mCAAmC;AAAA,IACjC,YAAY;AAAA,IACZ,aAAa;AAAA,IACb,SAAS;AAAA,EACX;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,uCAAuC,EAAE,YAAY,OAAO,aAAa,MAAM,SAAS,+BAA+B;AAAA,EACvH,yCAAyC,EAAE,YAAY,QAAQ,aAAa,QAAQ,SAAS,+BAA+B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAO5H,0BAA0B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,6BAA6B;AAAA,EACvG,8BAA8B,EAAE,YAAY,MAAM,aAAa,MAAM,SAAS,6BAA6B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAW3G,2BAA2B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,2BAA2B;AAAA;AAAA,EAE1H,gCAAgC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,2BAA2B;AAAA;AAAA;AAAA;AAAA;AAAA,EAM/H,2BAA2B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,2BAA2B;AAAA,EAC1H,gCAAgC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,2BAA2B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAU/H,gCAAgC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACvG,8BAA8B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACrG,iCAAiC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACxG,8BAA8B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACrG,4BAA4B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,GAAK,SAAS,qBAAqB;AAAA,EACrH,gCAAgC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACvG,8BAA8B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACrG,+BAA+B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACtG,+BAA+B,EAAE,YAAY,MAAM,gBAAgB,OAAO,aAAa,MAAM,SAAS,GAAG;AAAA,EACzG,gCAAgC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACvG,mCAAmC,EAAE,YAAY,GAAK,gBAAgB,KAAK,aAAa,GAAK,SAAS,GAAG;AAAA,EACzG,kCAAkC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACzG,2BAA2B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,GAAK,SAAS,GAAG;AAAA,EAClG,4BAA4B,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,KAAK,SAAS,GAAG;AAAA,EACnG,6BAA6B,EAAE,YAAY,MAAM,gBAAgB,OAAO,aAAa,MAAM,SAAS,GAAG;AAAA;AAAA,EAEvG,qCAAqC,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,GAAG,SAAS,GAAG;AAAA;AAAA,EAE1G,yBAAyB,EAAE,YAAY,KAAK,gBAAgB,MAAM,aAAa,GAAG,SAAS,GAAG;AAAA,EAC9F,2BAA2B,EAAE,YAAY,MAAM,gBAAgB,OAAO,aAAa,GAAG,SAAS,GAAG;AACpG;AAEO,SAAS,SAAS,UAAkB,OAAyC;AAClF,QAAM,QAAQ,QAAQ,GAAG,QAAQ,IAAI,KAAK,EAAE;AAC5C,MAAI,MAAO,QAAO;AAKlB,QAAM,OAAO,MAAM,QAAQ,WAAW,EAAE;AACxC,MAAI,SAAS,MAAO,QAAO,QAAQ,GAAG,QAAQ,IAAI,IAAI,EAAE;AACxD,SAAO;AACT;","names":[]}

package/dist/index.d.ts

@@ -130,6 +130,11 @@ interface ChatRequest {
      *  (components' requirement while building @broberg/chat, 2026-08-27, where two
      *  knowledge bases are written for readers with different permissions.) */
     promptCacheKey?: string;
+    /** F039.2 — prompt caching is ON by default where the provider supports it.
+     *  Set false to opt out (e.g. a one-shot call whose prefix will never repeat).
+     *  Off has no cost benefit and no penalty; a cache MISS was measured to cost
+     *  exactly the same as a call with no key at all. */
+    promptCache?: boolean;
     temperature?: number;
     /** "json" → request JSON-object output where the provider supports it (F009). */
     responseFormat?: "json" | "text";
@@ -546,7 +551,6 @@ declare const messageSchema: z.ZodObject<{
     }>, "many">>;
     toolCallId: z.ZodOptional<z.ZodString>;
 }, "strip", z.ZodTypeAny, {
-    role: "system" | "user" | "assistant" | "tool";
     content: string | ({
         text: string;
         type: "text";
@@ -555,6 +559,7 @@ declare const messageSchema: z.ZodObject<{
         type: "image";
         mimeType?: string | undefined;
     })[];
+    role: "system" | "user" | "assistant" | "tool";
     toolCalls?: {
         id: string;
         name: string;
@@ -562,7 +567,6 @@ declare const messageSchema: z.ZodObject<{
     }[] | undefined;
     toolCallId?: string | undefined;
 }, {
-    role: "system" | "user" | "assistant" | "tool";
     content: string | ({
         text: string;
         type: "text";
@@ -571,6 +575,7 @@ declare const messageSchema: z.ZodObject<{
         type: "image";
         mimeType?: string | undefined;
     })[];
+    role: "system" | "user" | "assistant" | "tool";
     toolCalls?: {
         id: string;
         name: string;
@@ -613,6 +618,9 @@ declare const chatInputSchema: z.ZodObject<{
      *  Derive it from (tenant, conversation): the key is a shared-prefix identity,
      *  so two tenants with identical transcripts must get different keys. */
     promptCacheKey: z.ZodOptional<z.ZodString>;
+    /** F039.2 — prompt caching is ON by default on providers that support it.
+     *  false opts out; an explicit promptCacheKey always wins. */
+    promptCache: z.ZodOptional<z.ZodBoolean>;
     prompt: z.ZodOptional<z.ZodString>;
     messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
         role: z.ZodEnum<["system", "user", "assistant", "tool"]>;
@@ -653,7 +661,6 @@ declare const chatInputSchema: z.ZodObject<{
         }>, "many">>;
         toolCallId: z.ZodOptional<z.ZodString>;
     }, "strip", z.ZodTypeAny, {
-        role: "system" | "user" | "assistant" | "tool";
         content: string | ({
             text: string;
             type: "text";
@@ -662,6 +669,7 @@ declare const chatInputSchema: z.ZodObject<{
             type: "image";
             mimeType?: string | undefined;
         })[];
+        role: "system" | "user" | "assistant" | "tool";
         toolCalls?: {
             id: string;
             name: string;
@@ -669,7 +677,6 @@ declare const chatInputSchema: z.ZodObject<{
         }[] | undefined;
         toolCallId?: string | undefined;
     }, {
-        role: "system" | "user" | "assistant" | "tool";
         content: string | ({
             text: string;
             type: "text";
@@ -678,6 +685,7 @@ declare const chatInputSchema: z.ZodObject<{
             type: "image";
             mimeType?: string | undefined;
         })[];
+        role: "system" | "user" | "assistant" | "tool";
         toolCalls?: {
             id: string;
             name: string;
@@ -708,7 +716,6 @@ declare const chatInputSchema: z.ZodObject<{
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     prompt?: string | undefined;
     messages?: {
-        role: "system" | "user" | "assistant" | "tool";
         content: string | ({
             text: string;
             type: "text";
@@ -717,6 +724,7 @@ declare const chatInputSchema: z.ZodObject<{
             type: "image";
             mimeType?: string | undefined;
         })[];
+        role: "system" | "user" | "assistant" | "tool";
         toolCalls?: {
             id: string;
             name: string;
@@ -732,6 +740,7 @@ declare const chatInputSchema: z.ZodObject<{
     temperature?: number | undefined;
     purpose?: string | undefined;
     promptCacheKey?: string | undefined;
+    promptCache?: boolean | undefined;
     maxTokens?: number | undefined;
     responseFormat?: "text" | "json" | undefined;
     override?: {
@@ -750,7 +759,6 @@ declare const chatInputSchema: z.ZodObject<{
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     prompt?: string | undefined;
     messages?: {
-        role: "system" | "user" | "assistant" | "tool";
         content: string | ({
             text: string;
             type: "text";
@@ -759,6 +767,7 @@ declare const chatInputSchema: z.ZodObject<{
             type: "image";
             mimeType?: string | undefined;
         })[];
+        role: "system" | "user" | "assistant" | "tool";
         toolCalls?: {
             id: string;
             name: string;
@@ -774,6 +783,7 @@ declare const chatInputSchema: z.ZodObject<{
     temperature?: number | undefined;
     purpose?: string | undefined;
     promptCacheKey?: string | undefined;
+    promptCache?: boolean | undefined;
     maxTokens?: number | undefined;
     responseFormat?: "text" | "json" | undefined;
     override?: {
@@ -1716,6 +1726,9 @@ declare const ttsInputSchema: z.ZodObject<{
     rate?: number | undefined;
 }>;
 declare const aiConfigSchema: z.ZodObject<{
+    /** F039.2 — client-wide opt-out from prompt caching (default: on where the
+     *  provider supports it). A per-call `promptCache` wins over this. */
+    promptCache: z.ZodOptional<z.ZodBoolean>;
     defaults: z.ZodOptional<z.ZodRecord<z.ZodEnum<["fast", "smart", "powerful", "cheap", "vision", "video", "embedding"]>, z.ZodObject<{
         provider: z.ZodString;
         model: z.ZodString;
@@ -1755,6 +1768,7 @@ declare const aiConfigSchema: z.ZodObject<{
         autoResolve?: boolean | undefined;
     }>>;
 }, "strip", z.ZodTypeAny, {
+    promptCache?: boolean | undefined;
     defaults?: Partial<Record<"fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding", {
         provider: string;
         model: string;
@@ -1771,6 +1785,7 @@ declare const aiConfigSchema: z.ZodObject<{
         autoResolve?: boolean | undefined;
     } | undefined;
 }, {
+    promptCache?: boolean | undefined;
     defaults?: Partial<Record<"fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding", {
         provider: string;
         model: string;
@@ -2060,6 +2075,13 @@ interface OpenAICompatibleConfig {
      *  response's `usage.cost` (USD) as costUsd, falling back to the pricing table.
      *  Only OpenRouter returns this field — openai/deepinfra leave it false. */
     costFromResponseField?: boolean;
+    /** F039.2 — this provider accepts `prompt_cache_key` and caches a shared prefix
+     *  on it. Mistral only, deliberately: an unknown field is ignored by some
+     *  OpenAI-compatible servers and rejected with a 400 by others, so sending it
+     *  everywhere would trade a saving for an outage. Providers that cache
+     *  AUTOMATICALLY (openai, deepseek, gemini) need no key — they only need their
+     *  cached counts read back, which happens for every provider below. */
+    supportsPromptCacheKey?: boolean;
 }
 declare function makeOpenAICompatibleAdapter(config: OpenAICompatibleConfig): ProviderAdapter;
 
@@ -2081,8 +2103,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.30.1";
-declare const SDK_TAG: "@broberg/ai-sdk@0.30.1";
+declare const VERSION: "0.32.0";
+declare const SDK_TAG: "@broberg/ai-sdk@0.32.0";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.

package/dist/index.js

@@ -10,7 +10,7 @@ import {
 } from "./chunk-V2PD522L.js";
 import {
   getPrice
-} from "./chunk-LKVCPMVI.js";
+} from "./chunk-DEIY7O3T.js";
 
 // src/transport/http.ts
 async function httpTransport(req) {
@@ -485,6 +485,16 @@ function mapAnthropicStop(reason) {
 }
 
 // src/providers/openai-compatible.ts
+function autoCacheKey(messages) {
+  const system = messages.find((m) => m.role === "system");
+  if (!system || typeof system.content !== "string" || system.content.length < 200) return void 0;
+  let h = 2166136261;
+  for (let i = 0; i < system.content.length; i++) {
+    h ^= system.content.charCodeAt(i);
+    h = Math.imul(h, 16777619) >>> 0;
+  }
+  return `ai-sdk-auto-${h.toString(16)}-${system.content.length}`;
+}
 function toOpenAIMessage(m) {
   if (typeof m.content === "string") {
     const base = { role: m.role, content: m.content };
@@ -523,7 +533,10 @@ function makeOpenAICompatibleAdapter(config) {
     if (req.maxTokens !== void 0) body.max_tokens = req.maxTokens;
     if (req.temperature !== void 0) body.temperature = req.temperature;
     if (req.responseFormat === "json") body.response_format = { type: "json_object" };
-    if (req.promptCacheKey !== void 0) body.prompt_cache_key = req.promptCacheKey;
+    if (config.supportsPromptCacheKey && req.promptCache !== false) {
+      const k = req.promptCacheKey ?? autoCacheKey(req.messages);
+      if (k !== void 0) body.prompt_cache_key = k;
+    }
     if (config.costFromResponseField) body.usage = { include: true };
     const res = await httpTransport({
       spec: req.spec,
@@ -794,6 +807,12 @@ var GEMINI_IMAGE_PRICE_PER_IMAGE = {
   "gemini-3-pro-image-preview": 0.134
   // was $0.039 — wrong (that's the flash price); pro is $0.134
 };
+function splitCached(meta) {
+  const prompt = meta?.promptTokenCount ?? 0;
+  const cached = meta?.cachedContentTokenCount;
+  if (cached === void 0) return { inputTokens: prompt };
+  return { inputTokens: Math.max(0, prompt - cached), cacheReadTokens: cached };
+}
 function partsFrom(content) {
   if (typeof content === "string") return [{ text: content }];
   return content.map((p) => {
@@ -868,7 +887,7 @@ function geminiAdapter(config = {}) {
       model: req.spec.model,
       transport: "http",
       capability: "chat",
-      inputTokens: data.usageMetadata?.promptTokenCount ?? 0,
+      ...splitCached(data.usageMetadata),
       outputTokens: data.usageMetadata?.candidatesTokenCount ?? 0
     });
     const result = { text, usage };
@@ -890,6 +909,7 @@ function geminiAdapter(config = {}) {
     const toolCalls = [];
     let inputTokens = 0;
     let outputTokens = 0;
+    let cacheReadTokens;
     let finishReason = null;
     for await (const data of stream) {
       let chunk;
@@ -908,7 +928,9 @@ function geminiAdapter(config = {}) {
       }
       if (candidate?.finishReason) finishReason = candidate.finishReason;
       if (chunk.usageMetadata) {
-        inputTokens = chunk.usageMetadata.promptTokenCount ?? inputTokens;
+        const split = splitCached(chunk.usageMetadata);
+        inputTokens = chunk.usageMetadata.promptTokenCount === void 0 ? inputTokens : split.inputTokens;
+        if (split.cacheReadTokens !== void 0) cacheReadTokens = split.cacheReadTokens;
         outputTokens = chunk.usageMetadata.candidatesTokenCount ?? outputTokens;
       }
     }
@@ -921,7 +943,8 @@ function geminiAdapter(config = {}) {
       transport: "http",
       capability: "chat",
       inputTokens,
-      outputTokens
+      outputTokens,
+      ...cacheReadTokens === void 0 ? {} : { cacheReadTokens }
     });
     yield { type: "usage", costUsd: usage.costUsd, model: usage.model, usage };
     yield {
@@ -965,7 +988,7 @@ function geminiAdapter(config = {}) {
       model: req.spec.model,
       transport: "http",
       capability: "image",
-      inputTokens: data.usageMetadata?.promptTokenCount ?? 0,
+      ...splitCached(data.usageMetadata),
       outputTokens: data.usageMetadata?.candidatesTokenCount ?? 0
     });
     usage.costUsd = config.pricePerImage ?? GEMINI_IMAGE_PRICE_PER_IMAGE[req.spec.model] ?? 0;
@@ -1150,7 +1173,7 @@ var VOXTRAL_PRICE_PER_MIN = {
 };
 function mistralAdapter(config = {}) {
   const baseUrl = config.baseUrl ?? "https://api.mistral.ai/v1";
-  const base = makeOpenAICompatibleAdapter({ name: "mistral", baseUrl, apiKey: config.apiKey });
+  const base = makeOpenAICompatibleAdapter({ name: "mistral", baseUrl, apiKey: config.apiKey, supportsPromptCacheKey: true });
   function key() {
     const k = config.apiKey ?? process.env.MISTRAL_API_KEY;
     if (!k) throw new Error("mistral adapter: API key not set (env MISTRAL_API_KEY)");
@@ -1726,7 +1749,7 @@ function vertexAdapter(config = {}) {
     }
     const data = await res.json();
     const text = (data.candidates?.[0]?.content?.parts ?? []).map((p) => p.text ?? "").join("").trim();
-    const inputTokens = data.usageMetadata?.promptTokenCount ?? 0;
+    const { inputTokens, cacheReadTokens } = splitCached(data.usageMetadata);
     const outputTokens = data.usageMetadata?.candidatesTokenCount ?? 0;
     const usage = freshUsage({
       provider: "vertex",
@@ -1734,9 +1757,10 @@ function vertexAdapter(config = {}) {
       transport: "http",
       capability: "vision",
       inputTokens,
-      outputTokens
+      outputTokens,
+      ...cacheReadTokens === void 0 ? {} : { cacheReadTokens }
     });
-    usage.costUsd = computeCost("vertex", req.spec.model, inputTokens, outputTokens);
+    usage.costUsd = computeCost("vertex", req.spec.model, inputTokens, outputTokens, cacheReadTokens ?? 0);
     return { text, usage };
   }
   return { name: "vertex", animate, vision };
@@ -2631,6 +2655,9 @@ var chatInputSchema = z.object({
    *  Derive it from (tenant, conversation): the key is a shared-prefix identity,
    *  so two tenants with identical transcripts must get different keys. */
   promptCacheKey: z.string().optional(),
+  /** F039.2 — prompt caching is ON by default on providers that support it.
+   *  false opts out; an explicit promptCacheKey always wins. */
+  promptCache: z.boolean().optional(),
   prompt: z.string().optional(),
   messages: z.array(messageSchema).optional(),
   system: z.string().optional(),
@@ -2779,6 +2806,9 @@ var availabilitySchema = z.object({
   fallback: z.union([z.string(), z.array(z.string())]).optional()
 });
 var aiConfigSchema = z.object({
+  /** F039.2 — client-wide opt-out from prompt caching (default: on where the
+   *  provider supports it). A per-call `promptCache` wins over this. */
+  promptCache: z.boolean().optional(),
   defaults: z.record(tierSchema, tierSpecSchema).optional(),
   // Functions can't be deeply validated — z.custom asserts the TS type and
   // passes the value through untouched.
@@ -2792,8 +2822,8 @@ var aiConfigSchema = z.object({
 });
 
 // src/version.ts
-var VERSION = "0.30.1";
-var SDK_TAG = "@broberg/ai-sdk@0.30.1";
+var VERSION = "0.32.0";
+var SDK_TAG = "@broberg/ai-sdk@0.32.0";
 
 // src/cost/sinks/upmetrics.ts
 function upmetricsSink(config) {
@@ -3077,7 +3107,7 @@ function createAI(config = {}) {
         invoke: async (spec) => {
           const adapter = pickProvider(spec.provider);
           if (!adapter.chat) throw new Error(`createAI: provider "${spec.provider}" does not support chat`);
-          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, responseFormat: input.responseFormat, promptCacheKey: input.promptCacheKey });
+          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, responseFormat: input.responseFormat, promptCacheKey: input.promptCacheKey, promptCache: input.promptCache ?? cfg.promptCache });
         }
       });
     },