@broberg/ai-sdk 0.55.3 → 0.56.1

package/dist/{chunk-ILZ33HZO.js → chunk-BOZYLPIM.js}

@@ -1,213 +1,7 @@
-// src/cost/pricing.ts
-var V = "2026-06-02";
-var MS = "2026-06-04-mistral.ai";
-var AN = "2026-10-08-platform.claude.com";
-var PRICING = {
-  // Anthropic (direct API). DEFAULT_TIER_MAP: fast/cheap=haiku, smart/vision=sonnet, powerful=opus.
-  // $1.00/$5.00 — Anthropic's published Haiku 4.5 rate, confirmed against the
-  // claude-api skill's model table 2026-09-03. F048: this row said $0.80/$4.00 while
-  // "openrouter:anthropic/claude-haiku-4.5" said $1.00/$5.00 — the SAME model at two
-  // prices, and the shorter, likelier id carried the wrong one. A consumer pricing
-  // 68M haiku tokens (upmetrics' real volume) was told 20% under.
-  "anthropic:claude-haiku-4-5": {
-    inputPer1M: 1,
-    outputPer1M: 5,
-    cacheReadPer1M: 0.1,
-    cacheWritePer1M: 1.25,
-    version: "2026-09-03"
-  },
-  "anthropic:claude-sonnet-4-6": {
-    inputPer1M: 3,
-    outputPer1M: 15,
-    cacheReadPer1M: 0.3,
-    cacheWritePer1M: 3.75,
-    version: V
-  },
-  // $5.00/$25.00 — Anthropic's published Opus 4.8 rate, confirmed against the
-  // claude-api skill's model table 2026-09-03.
-  //
-  // FOUND BY THE CONTRADICTION GUARD ON ITS FIRST RUN, not by a report. This row said
-  // $15.00/$75.00 while "openrouter:anthropic/claude-opus-4.8" said $5.00/$25.00 —
-  // THREE TIMES too high, the opposite direction from the haiku error and a larger
-  // multiple. Nobody had noticed, because both numbers looked like answers.
-  "anthropic:claude-opus-4-8": {
-    inputPer1M: 5,
-    outputPer1M: 25,
-    cacheReadPer1M: 0.5,
-    cacheWritePer1M: 6.25,
-    version: "2026-09-03"
-  },
-  // F075 — the rest of what Anthropic's /v1/models lists for our key, read 2026-10-08
-  // from platform.claude.com/docs/en/about-claude/pricing. Cache-read is NOT 0.1x
-  // everywhere: Opus/Sonnet 5.5 are 0.05x and Fable 5.1 is 0.025x. cacheWrite is the
-  // 5-minute rate (the only one the SDK writes).
-  "anthropic:claude-fable-5-1": { inputPer1M: 10, outputPer1M: 50, cacheReadPer1M: 0.25, cacheWritePer1M: 12.5, version: AN },
-  "anthropic:claude-fable-5": { inputPer1M: 10, outputPer1M: 50, cacheReadPer1M: 1, cacheWritePer1M: 12.5, version: AN },
-  "anthropic:claude-opus-5-5": { inputPer1M: 4, outputPer1M: 20, cacheReadPer1M: 0.2, cacheWritePer1M: 5, version: AN },
-  "anthropic:claude-opus-5": { inputPer1M: 5, outputPer1M: 25, cacheReadPer1M: 0.5, cacheWritePer1M: 6.25, version: AN },
-  "anthropic:claude-opus-4-7": { inputPer1M: 5, outputPer1M: 25, cacheReadPer1M: 0.5, cacheWritePer1M: 6.25, version: AN },
-  "anthropic:claude-opus-4-6": { inputPer1M: 5, outputPer1M: 25, cacheReadPer1M: 0.5, cacheWritePer1M: 6.25, version: AN },
-  "anthropic:claude-opus-4-5": { inputPer1M: 5, outputPer1M: 25, cacheReadPer1M: 0.5, cacheWritePer1M: 6.25, version: AN },
-  "anthropic:claude-sonnet-5-5": { inputPer1M: 2, outputPer1M: 10, cacheReadPer1M: 0.1, cacheWritePer1M: 2.5, version: AN },
-  "anthropic:claude-sonnet-5": { inputPer1M: 2, outputPer1M: 10, cacheReadPer1M: 0.2, cacheWritePer1M: 2.5, version: AN },
-  "anthropic:claude-sonnet-4-5": { inputPer1M: 3, outputPer1M: 15, cacheReadPer1M: 0.3, cacheWritePer1M: 3.75, version: AN },
-  "anthropic:claude-haiku-5-5": {
-    inputPer1M: 0.1,
-    outputPer1M: 0.5,
-    cacheReadPer1M: 0.01,
-    cacheWritePer1M: 0.125,
-    version: AN,
-    longPrompt: { aboveTokens: 1e5, inputPer1M: 0.5, outputPer1M: 2.5, cacheReadPer1M: 0.05, cacheWritePer1M: 0.625 }
-  },
-  // OpenAI. embedding default tier = text-embedding-3-small (no output tokens).
-  // Cached input is 50% of the input rate — $1.25 vs $2.50 (gpt-4o) and $0.075 vs
-  // $0.15 (gpt-4o-mini), read from developers.openai.com/api/docs/pricing on
-  // 2026-08-27. NOTE IT IS 50%, NOT the 10% Mistral and Gemini charge: assuming one
-  // uniform discount across providers would have understated OpenAI's cost by 4x on
-  // the cached half. Caching is automatic above 1,024 tokens — no key, no opt-in.
-  // The embedding models list NO cached price (embeddings do not cache), so they
-  // deliberately get no row here rather than a guessed one.
-  "openai:text-embedding-3-small": { inputPer1M: 0.02, outputPer1M: 0, version: V },
-  "openai:text-embedding-3-large": { inputPer1M: 0.13, outputPer1M: 0, version: V },
-  "openai:gpt-4o": { inputPer1M: 2.5, cacheReadPer1M: 1.25, outputPer1M: 10, version: "2026-08-27-developers.openai.com" },
-  "openai:gpt-4o-mini": { inputPer1M: 0.15, cacheReadPer1M: 0.075, outputPer1M: 0.6, version: "2026-08-27-developers.openai.com" },
-  // Whisper is priced per MINUTE — see MEDIA_PRICING in ./media-pricing.ts. It used to
-  // sit here as 0/0 "so token-based compute never charges it", and inwardly that worked.
-  // Outwardly it did not: getModelPrice("whisper-1") answered `{unit:"per_1m_tokens",
-  // inputPer1M: 0}` — i.e. "this model is free" — because media rows are consulted LAST
-  // and a token row was already there. The per-minute price was unreachable by its own
-  // id. computeCost still returns 0 for it (no entry) and transcribe still computes its
-  // own cost, so nothing about billing changed; only the public answer stopped lying.
-  // OpenRouter (meta-router — model slugs include the upstream vendor). Slugs use
-  // dots (claude-sonnet-4.6) to match OpenRouter's live ids; the dashed forms
-  // never matched a real call. Caught by the F014 catalogue research.
-  "openrouter:anthropic/claude-sonnet-4.6": { inputPer1M: 3, outputPer1M: 15, version: V },
-  // OpenRouter ground-truth $1/$5 — a markup over Anthropic-direct's $0.8/$4
-  // (the `anthropic:` entry above). Was masked while the slug used dashes.
-  "openrouter:anthropic/claude-haiku-4.5": { inputPer1M: 1, outputPer1M: 5, version: "2026-06-04" },
-  "openrouter:google/gemini-2.5-flash": { inputPer1M: 0.3, outputPer1M: 2.5, version: V },
-  // Ground-truth from OpenRouter /api/v1/models, re-read 2026-10-02 (F067.5; was 0.279/1.2).
-  "openrouter:minimax/minimax-m2.7": {
-    inputPer1M: 0.21,
-    outputPer1M: 0.84,
-    version: "2026-10-02-openrouter-live"
-  },
-  // DeepSeek V4 (CN-hosted — NOT GDPR-safe; non-personal-data workloads only).
-  // On 2026-05-22 DeepSeek made the "75% off" promo the permanent official price.
-  // V4-Pro $0.435/$0.87 is ~34x cheaper than GPT-5.5 on output; flash is cheaper
-  // still. Numbers match OpenRouter /api/v1/models 1:1 (no router markup). A strong
-  // cheap route for fleet background work once `claude -p` is API-billed (15 Jun).
-  // Re-read from OpenRouter /api/v1/models 2026-10-02 (F067.5). These moved three
-  // times in a month (Sep 3: pro 1.02, Oct 1: flash 0.042, Oct 2: flash 0.028) — an
-  // OpenRouter call books usage.cost from the response anyway; this row is the fallback.
-  "openrouter:deepseek/deepseek-v4-pro": { inputPer1M: 0.2088, outputPer1M: 0.4176, version: "2026-10-02-openrouter-live" },
-  "openrouter:deepseek/deepseek-v4-flash": { inputPer1M: 0.028, outputPer1M: 0.056, version: "2026-10-02-openrouter-live" },
-  // DeepSeek DIRECT API (provider "deepseek", F030 non-PII secondary). Rates from
-  // api-docs.deepseek.com 2026-06-30 ($0.14/$0.28 per 1M; both map to deepseek-v4-flash).
-  // `deepseek-chat` (non-thinking) + `deepseek-reasoner` (thinking) DEPRECATE 2026-07-24.
-  // (The bare `deepseek-v4-flash` basename is already priced via the openrouter entry
-  // above — kept distinct here to avoid a basename collision in the F027 pricing-API.)
-  // 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" },
-  // F066 — TypeSafe Jev. docs.typesafe.ai/models, read 2026-10-01: "$42 per billion
-  // tokens (input only)", output free. Priced on the VERSIONED id because the response
-  // reports the model that actually ran ("jev-1.13.0") even when you ask for
-  // "jev-latest" — so a future version arrives unpriced and is booked as such, rather
-  // than inheriting a number nobody checked for it.
-  "typesafe:jev-1.13.0": { inputPer1M: 0.042, outputPer1M: 0, version: "2026-10-01-typesafe-official" },
-  // 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, 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, 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, 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).
-  // Mistral (direct, La Plateforme). Official prices from mistral.ai/pricing
-  // (2026-06-04, per Christian's CD report). EU/Paris-hosted — the designated
-  // GDPR-safe provider for client/personal-data workloads (see F015). NB:
-  // medium-3.5 is the premium "Vibe" coding tier ($1.5/$7.5); Large 3 ($0.5/$1.5)
-  // is the cheaper frontier general-purpose model despite the higher number.
-  "mistral:mistral-large-latest": { inputPer1M: 0.5, cacheReadPer1M: 0.05, outputPer1M: 1.5, version: MS },
-  "mistral:mistral-large-2512": { inputPer1M: 0.5, cacheReadPer1M: 0.05, outputPer1M: 1.5, version: MS },
-  // Large 4 (F074.1), measured 2026-10-08 on docs.mistral.ai/models/mistral-large-4:
-  // LIST price. The page also shows a 50% "Public Preview" sale ($0.68/$0.07/$2.09)
-  // with no end date; booking the sale would under-report the day it ends, unseen.
-  // Cached input shows as "$0.14": that is 10% of $1.36 rounded to cents (the sale's
-  // "$0.07" is 0.068 the same way), and 10% is the rate measured for Mistral (F039).
-  "mistral:mistral-large-4": { inputPer1M: 1.36, cacheReadPer1M: 0.136, outputPer1M: 4.18, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-large-4-0": { inputPer1M: 1.36, cacheReadPer1M: 0.136, outputPer1M: 4.18, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-medium-latest": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: MS },
-  "mistral:mistral-medium-3.5": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: MS },
-  // F076.2 — Mistral's model API now lists mistral-medium-3 as an alias of Medium 3.5
-  // (measured 2026-10-08), so it bills at 3.5's price, not Medium 3's $0.40/$2.00.
-  "mistral:mistral-medium-3": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  // F076.1 — Small 4 (2603) is $0.15/$0.60, measured 2026-10-08 on
-  // docs.mistral.ai/models/mistral-small-4-0-26-03 AND OpenRouter's mistral-small-2603.
-  // This row said $0.10/$0.30 (Small 3's price, carried over when -latest moved to
-  // 2603), so the fleet's fast/cheap tiers were booked at about half their cost.
-  "mistral:mistral-small-latest": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.6, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-small-2603": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.6, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:ministral-3b-latest": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.1, version: MS },
-  "mistral:ministral-8b-latest": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.15, version: MS },
-  "mistral:ministral-14b-latest": { inputPer1M: 0.2, cacheReadPer1M: 0.02, outputPer1M: 0.2, version: MS },
-  // F076.2 — both Magistral aliases now point at the unified models (measured
-  // 2026-10-08 via /v1/models): magistral-medium-latest → Medium 3.5, magistral-small-
-  // latest → Small 4. They bill at those models' prices, not the old Magistral rates.
-  "mistral:magistral-medium-latest": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:magistral-small-latest": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.6, version: "2026-10-08-docs.mistral.ai" },
-  // F076.2 — ids Mistral's model API lists for our key that had no row (a call to any
-  // of them was booked at $0). Prices from docs.mistral.ai, 2026-10-08; cache-read at
-  // the measured 10% (F039).
-  "mistral:mistral-vibe-cli-fast": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.6, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-medium-3-5": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-medium": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-medium-2604": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-vibe-cli-latest": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-vibe-cli-with-tools": { inputPer1M: 1.5, cacheReadPer1M: 0.15, outputPer1M: 7.5, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:ministral-14b-2512": { inputPer1M: 0.2, cacheReadPer1M: 0.02, outputPer1M: 0.2, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:ministral-8b-2512": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.15, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:ministral-3b-2512": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.1, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:voxtral-small-2507": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:voxtral-small-latest": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0.4, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:codestral-2508": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 0.9, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-code-latest": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 0.9, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:mistral-code-fim-latest": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 0.9, version: "2026-10-08-docs.mistral.ai" },
-  "mistral:devstral-latest": { inputPer1M: 0.4, cacheReadPer1M: 0.04, outputPer1M: 2, version: MS },
-  "mistral:codestral-latest": { inputPer1M: 0.3, cacheReadPer1M: 0.03, outputPer1M: 0.9, version: MS },
-  "mistral:open-mistral-nemo": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0.15, version: MS },
-  // Moderation (F016.4) — per input token; output 0. (OCR is per-page in the adapter.)
-  "mistral:mistral-moderation-latest": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0, version: MS },
-  // Embeddings (F016.5) — per input token.
-  "mistral:mistral-embed": { inputPer1M: 0.1, cacheReadPer1M: 0.01, outputPer1M: 0, version: MS },
-  "mistral:codestral-embed": { inputPer1M: 0.15, cacheReadPer1M: 0.015, outputPer1M: 0, version: MS },
-  // F048 — reported by upmetrics as an unknown model (6 calls). Official mistral.ai/pricing.
-  "mistral:pixtral-large-latest": { inputPer1M: 2, cacheReadPer1M: 0.2, outputPer1M: 6, version: "2026-09-03" }
-};
-function stripDatedSuffix(model) {
-  return model.replace(/-\d{8}$/, "");
-}
-function getPrice(provider, model) {
-  const exact = PRICING[`${provider}:${model}`];
-  if (exact) return exact;
-  const base = stripDatedSuffix(model);
-  if (base !== model) return PRICING[`${provider}:${base}`];
-  return void 0;
-}
+import {
+  PRICING,
+  stripDatedSuffix
+} from "./chunk-QVMGFHCF.js";
 
 // src/cost/media-pricing.ts
 var MEDIA_PRICING_CHECKED_AT = "2026-09-05";
@@ -516,8 +310,6 @@ function resetPricingWarningForTests() {
 }
 
 export {
-  PRICING,
-  getPrice,
   MEDIA_PRICING_CHECKED_AT,
   MEDIA_PRICING,
   DEFAULT_CLIP_SEC,
@@ -533,4 +325,4 @@ export {
   warnIfPricingStale,
   resetPricingWarningForTests
 };
-//# sourceMappingURL=chunk-ILZ33HZO.js.map
+//# sourceMappingURL=chunk-BOZYLPIM.js.map