@bymax-one/nest-ai-tokens 1.0.0

package/dist/prices/index.cjs

@@ -0,0 +1,218 @@
+'use strict';
+
+// src/prices/model-prices.seed.ts
+var ZERO_RATES = {
+  inputNanoUsdPerMillion: 0n,
+  outputNanoUsdPerMillion: 0n,
+  cacheReadNanoUsdPerMillion: 0n,
+  cacheWrite5mNanoUsdPerMillion: 0n,
+  cacheWrite1hNanoUsdPerMillion: 0n,
+  reasoningNanoUsdPerMillion: 0n,
+  audioInNanoUsdPerMillion: 0n,
+  audioOutNanoUsdPerMillion: 0n,
+  imageInNanoUsdPerMillion: 0n,
+  imageOutNanoUsdPerMillion: 0n
+};
+function makeRow(input) {
+  return {
+    provider: input.provider,
+    model: input.model,
+    operation: input.operation ?? "chat",
+    serviceTier: input.serviceTier ?? "standard",
+    ...ZERO_RATES,
+    ...input.rates,
+    ...input.tierThresholdTokens !== void 0 ? { tierThresholdTokens: input.tierThresholdTokens } : {},
+    ...input.tierInputNanoUsdPerMillion !== void 0 ? { tierInputNanoUsdPerMillion: input.tierInputNanoUsdPerMillion } : {},
+    ...input.tierOutputNanoUsdPerMillion !== void 0 ? { tierOutputNanoUsdPerMillion: input.tierOutputNanoUsdPerMillion } : {},
+    ...input.unitRates !== void 0 ? { unitRates: input.unitRates } : {},
+    currency: "USD",
+    source: "snapshot"
+  };
+}
+var MODEL_PRICES_SEED = [
+  // ── OpenAI ───────────────────────────────────────────────────────────────
+  // source: OpenAI pricing snapshot 2026-07 — gpt-5 $1.25/$10 per M, cached input $0.125/M.
+  makeRow({
+    provider: "openai",
+    model: "gpt-5",
+    rates: {
+      inputNanoUsdPerMillion: 1250000000n,
+      outputNanoUsdPerMillion: 10000000000n,
+      cacheReadNanoUsdPerMillion: 125000000n,
+      reasoningNanoUsdPerMillion: 10000000000n
+    },
+    unitRates: { web_search_requests: 10000000n }
+    // VERIFY: $0.01 per built-in web-search call.
+  }),
+  // source: OpenAI pricing snapshot 2026-07 — Batch API ~50% of standard.
+  makeRow({
+    provider: "openai",
+    model: "gpt-5",
+    serviceTier: "batch",
+    rates: {
+      inputNanoUsdPerMillion: 625000000n,
+      outputNanoUsdPerMillion: 5000000000n,
+      cacheReadNanoUsdPerMillion: 62500000n,
+      reasoningNanoUsdPerMillion: 5000000000n
+    }
+  }),
+  // source: OpenAI pricing snapshot 2026-07 — gpt-5-mini $0.25/$2 per M.
+  makeRow({
+    provider: "openai",
+    model: "gpt-5-mini",
+    rates: {
+      inputNanoUsdPerMillion: 250000000n,
+      outputNanoUsdPerMillion: 2000000000n,
+      cacheReadNanoUsdPerMillion: 25000000n,
+      reasoningNanoUsdPerMillion: 2000000000n
+    }
+  }),
+  // source: OpenAI pricing snapshot 2026-07 — text-embedding-3-small $0.02/M.
+  makeRow({
+    provider: "openai",
+    model: "text-embedding-3-small",
+    operation: "embeddings",
+    rates: { inputNanoUsdPerMillion: 20000000n }
+  }),
+  // source: OpenAI pricing snapshot 2026-07 — text-embedding-3-large $0.13/M.
+  makeRow({
+    provider: "openai",
+    model: "text-embedding-3-large",
+    operation: "embeddings",
+    rates: { inputNanoUsdPerMillion: 130000000n }
+  }),
+  // ── Anthropic ────────────────────────────────────────────────────────────
+  // source: Anthropic pricing snapshot 2026-07 — Opus $15/$75 per M; cache read 0.1×, write5m 1.25×, write1h 2× input.
+  makeRow({
+    provider: "anthropic",
+    model: "claude-opus-4",
+    rates: {
+      inputNanoUsdPerMillion: 15000000000n,
+      outputNanoUsdPerMillion: 75000000000n,
+      cacheReadNanoUsdPerMillion: 1500000000n,
+      cacheWrite5mNanoUsdPerMillion: 18750000000n,
+      cacheWrite1hNanoUsdPerMillion: 30000000000n,
+      reasoningNanoUsdPerMillion: 75000000000n
+    },
+    unitRates: { web_search_requests: 10000000n }
+    // source: Anthropic web search $10 / 1,000 requests.
+  }),
+  // source: Anthropic pricing snapshot 2026-07 — Sonnet $3/$15 per M.
+  makeRow({
+    provider: "anthropic",
+    model: "claude-sonnet-4",
+    rates: {
+      inputNanoUsdPerMillion: 3000000000n,
+      outputNanoUsdPerMillion: 15000000000n,
+      cacheReadNanoUsdPerMillion: 300000000n,
+      cacheWrite5mNanoUsdPerMillion: 3750000000n,
+      cacheWrite1hNanoUsdPerMillion: 6000000000n,
+      reasoningNanoUsdPerMillion: 15000000000n
+    }
+  }),
+  // source: Anthropic pricing snapshot 2026-07 — Haiku $0.80/$4 per M.
+  makeRow({
+    provider: "anthropic",
+    model: "claude-haiku-4",
+    rates: {
+      inputNanoUsdPerMillion: 800000000n,
+      outputNanoUsdPerMillion: 4000000000n,
+      cacheReadNanoUsdPerMillion: 80000000n,
+      cacheWrite5mNanoUsdPerMillion: 1000000000n,
+      cacheWrite1hNanoUsdPerMillion: 1600000000n,
+      reasoningNanoUsdPerMillion: 4000000000n
+    }
+  }),
+  // source: Anthropic pricing snapshot 2026-07 — Sonnet Batch API ~50% of standard.
+  makeRow({
+    provider: "anthropic",
+    model: "claude-sonnet-4",
+    serviceTier: "batch",
+    rates: {
+      inputNanoUsdPerMillion: 1500000000n,
+      outputNanoUsdPerMillion: 7500000000n,
+      cacheReadNanoUsdPerMillion: 150000000n,
+      cacheWrite5mNanoUsdPerMillion: 1875000000n,
+      cacheWrite1hNanoUsdPerMillion: 3000000000n,
+      reasoningNanoUsdPerMillion: 7500000000n
+    }
+  }),
+  // ── Google Gemini ────────────────────────────────────────────────────────
+  // source: Gemini pricing snapshot 2026-07 — 2.5 Pro $1.25/$10 (≤200k), $2.50/$15 above; cache read ~10% input.
+  makeRow({
+    provider: "gemini",
+    model: "gemini-2.5-pro",
+    rates: {
+      inputNanoUsdPerMillion: 1250000000n,
+      outputNanoUsdPerMillion: 10000000000n,
+      cacheReadNanoUsdPerMillion: 125000000n,
+      reasoningNanoUsdPerMillion: 10000000000n
+    },
+    tierThresholdTokens: 2e5,
+    tierInputNanoUsdPerMillion: 2500000000n,
+    tierOutputNanoUsdPerMillion: 15000000000n
+  }),
+  // source: Gemini pricing snapshot 2026-07 — 2.5 Flash $0.30/$2.50; long-context tier VERIFY.
+  makeRow({
+    provider: "gemini",
+    model: "gemini-2.5-flash",
+    rates: {
+      inputNanoUsdPerMillion: 300000000n,
+      outputNanoUsdPerMillion: 2500000000n,
+      cacheReadNanoUsdPerMillion: 30000000n,
+      reasoningNanoUsdPerMillion: 2500000000n
+    },
+    tierThresholdTokens: 2e5,
+    // VERIFY: Flash long-context surcharge.
+    tierInputNanoUsdPerMillion: 600000000n,
+    tierOutputNanoUsdPerMillion: 5000000000n
+  }),
+  // ── Mistral ──────────────────────────────────────────────────────────────
+  // source: Mistral pricing snapshot 2026-07 — Large $2/$6 per M.
+  makeRow({
+    provider: "mistral",
+    model: "mistral-large-latest",
+    rates: { inputNanoUsdPerMillion: 2000000000n, outputNanoUsdPerMillion: 6000000000n }
+  }),
+  // source: Mistral pricing snapshot 2026-07 — Medium $0.40/$2 per M.
+  makeRow({
+    provider: "mistral",
+    model: "mistral-medium-latest",
+    rates: { inputNanoUsdPerMillion: 400000000n, outputNanoUsdPerMillion: 2000000000n }
+  }),
+  // source: Mistral pricing snapshot 2026-07 — Small $0.10/$0.30 per M.
+  makeRow({
+    provider: "mistral",
+    model: "mistral-small-latest",
+    rates: { inputNanoUsdPerMillion: 100000000n, outputNanoUsdPerMillion: 300000000n }
+  }),
+  // ── DeepSeek / xAI / Groq (headline models) ──────────────────────────────
+  // source: DeepSeek pricing snapshot 2026-07 — chat $0.27/$1.10 per M; cache hit $0.07/M.
+  makeRow({
+    provider: "deepseek",
+    model: "deepseek-chat",
+    rates: {
+      inputNanoUsdPerMillion: 270000000n,
+      outputNanoUsdPerMillion: 1100000000n,
+      cacheReadNanoUsdPerMillion: 70000000n
+    }
+  }),
+  // source: xAI pricing snapshot 2026-07 — Grok $3/$15 per M.
+  makeRow({
+    provider: "xai",
+    model: "grok-4",
+    rates: {
+      inputNanoUsdPerMillion: 3000000000n,
+      outputNanoUsdPerMillion: 15000000000n,
+      reasoningNanoUsdPerMillion: 15000000000n
+    }
+  }),
+  // source: Groq pricing snapshot 2026-07 — Llama 3.3 70B $0.59/$0.79 per M. VERIFY.
+  makeRow({
+    provider: "groq",
+    model: "llama-3.3-70b-versatile",
+    rates: { inputNanoUsdPerMillion: 590000000n, outputNanoUsdPerMillion: 790000000n }
+  })
+];
+
+exports.MODEL_PRICES_SEED = MODEL_PRICES_SEED;

package/dist/prices/index.d.cts

@@ -0,0 +1,131 @@
+/**
+ * @fileoverview Catalog of AI operation kinds a usage record can describe. The
+ * operation is part of the price-resolution key; `'responses'` bills identically
+ * to `'chat'` and shares its price rows (see spec §5.1).
+ * @layer shared
+ */
+/** Every operation kind the library can meter and price. */
+declare const AI_OPERATIONS: readonly ["chat", "responses", "embeddings", "image", "video", "audio", "rerank", "moderation"];
+/** A logical AI operation kind. */
+type AiOperation = (typeof AI_OPERATIONS)[number];
+
+/**
+ * @fileoverview Catalog of response service tiers. Providers may silently
+ * downgrade a request (OpenAI returns `service_tier: "default"` when priority
+ * capacity is exhausted), so pricing keys off the tier reported in the RESPONSE,
+ * which is what the normalizers read (see spec §5.1).
+ * @layer shared
+ */
+/**
+ * Service tier of the actual response. `'batch'` is the Batch API (~50% discount);
+ * `'flex'` is OpenAI flex processing (batch rates, synchronous); `'priority'` is
+ * paid premium; `'standard'` is the default.
+ */
+declare const SERVICE_TIERS: readonly ["standard", "batch", "flex", "priority"];
+/** A response service tier. */
+type ServiceTier = (typeof SERVICE_TIERS)[number];
+
+/**
+ * @fileoverview Catalog of the AI providers this library recognizes out of the
+ * box. Custom OpenAI-compatible providers register their own id at runtime via
+ * `providerPresets.openaiCompatible(id)`, so the public {@link ProviderId} type
+ * (see `types/catalogs.ts`) widens this closed list with `(string & {})`.
+ * @layer shared
+ */
+/**
+ * The provider identifiers shipped with the library. Frozen tuple so the derived
+ * {@link KnownProviderId} union stays in sync with the runtime values.
+ */
+declare const PROVIDER_IDS: readonly ["openai", "azure-openai", "anthropic", "gemini", "vertex", "mistral", "bedrock", "openrouter", "deepseek", "xai", "groq"];
+/** A provider id known to the library at build time (excludes custom ids). */
+type KnownProviderId = (typeof PROVIDER_IDS)[number];
+
+/**
+ * @fileoverview Canonical catalog types that the rest of the library is generic
+ * over: provider/operation/tier unions, the metering scope, the rating mode, and
+ * the pure normalizer/preset contracts. Framework-free.
+ * @layer shared
+ */
+
+/**
+ * A provider identifier. Widened with `(string & {})` so custom
+ * OpenAI-compatible providers can register their own id while known ids keep
+ * editor autocomplete.
+ */
+type ProviderId = KnownProviderId | (string & {});
+
+/**
+ * @fileoverview The effective-dated price row and its insert type. Rates are
+ * integer nano-USD per 1,000,000 tokens (or per unit for non-token line items),
+ * so every cost computation is exact integer math.
+ * @layer shared
+ */
+
+/**
+ * One immutable, effective-dated price version for a (provider, model, operation,
+ * serviceTier) tuple. A call is rated at the row in effect at its timestamp.
+ */
+interface PriceVersion {
+    id: string;
+    provider: ProviderId;
+    model: string;
+    operation: AiOperation;
+    /** Part of the resolution key; `'standard'` is the default. */
+    serviceTier: ServiceTier;
+    /** Rates as integer nano-USD per 1,000,000 tokens. */
+    inputNanoUsdPerMillion: bigint;
+    outputNanoUsdPerMillion: bigint;
+    cacheReadNanoUsdPerMillion: bigint;
+    cacheWrite5mNanoUsdPerMillion: bigint;
+    cacheWrite1hNanoUsdPerMillion: bigint;
+    /** Usually equals output; kept separate for models that price reasoning differently. */
+    reasoningNanoUsdPerMillion: bigint;
+    audioInNanoUsdPerMillion: bigint;
+    audioOutNanoUsdPerMillion: bigint;
+    imageInNanoUsdPerMillion: bigint;
+    imageOutNanoUsdPerMillion: bigint;
+    /**
+     * Long-context tier threshold: when total input (all input-side categories)
+     * exceeds this, the tier rates replace the base input/output rates for the
+     * WHOLE call (all-or-nothing, matching Gemini's billing).
+     */
+    tierThresholdTokens?: number;
+    tierInputNanoUsdPerMillion?: bigint;
+    tierOutputNanoUsdPerMillion?: bigint;
+    /**
+     * Non-token line items in nano-USD PER UNIT, matched against
+     * `NormalizedUsage.serverToolUse` / `MeteringContext.extraUnits` by EXACT key —
+     * so the key must equal the count key the normalizer emits (e.g. Anthropic's
+     * `usage.server_tool_use.web_search_requests`) — e.g.
+     * `{ web_search_requests: 10_000_000n }` ($0.01 per web-search call).
+     */
+    unitRates?: Record<string, bigint>;
+    currency: 'USD';
+    effectiveFrom: Date;
+    /** `null` = current; set when a newer row supersedes it. */
+    effectiveTo: Date | null;
+    /** Provenance: `'snapshot' | 'manual' | 'import'`. */
+    source: string;
+}
+
+/**
+ * @fileoverview Pinned per-model price snapshot in bigint nano-USD per 1,000,000
+ * tokens (unit rates are per unit). Converted offline from published provider
+ * pricing (methodology mirrors LiteLLM's `model_prices_and_context_window.json`,
+ * see `scripts/convert-litellm-prices.mjs`). Snapshot date: 2026-07. Pinning
+ * keeps rates point-in-time stable; hosts add or override models via
+ * `PricingService.upsertPrice()`. Data only — the exported array carries no
+ * behavior. Each row's `source` is `'snapshot'`; rates flagged `VERIFY` should be
+ * spot-checked against the provider's pricing page during review.
+ * @layer prices
+ */
+
+/** A seed price row: a {@link PriceVersion} without the store-assigned identity/effective-date fields. */
+type SeedPriceRow = Omit<PriceVersion, 'id' | 'effectiveFrom' | 'effectiveTo'>;
+/**
+ * The pinned price snapshot. Rated at published standard-tier prices unless the
+ * row's `serviceTier` says otherwise; batch rows are ~50% of standard.
+ */
+declare const MODEL_PRICES_SEED: readonly SeedPriceRow[];
+
+export { MODEL_PRICES_SEED, type SeedPriceRow };

package/dist/prices/index.d.ts

@@ -0,0 +1,131 @@
+/**
+ * @fileoverview Catalog of AI operation kinds a usage record can describe. The
+ * operation is part of the price-resolution key; `'responses'` bills identically
+ * to `'chat'` and shares its price rows (see spec §5.1).
+ * @layer shared
+ */
+/** Every operation kind the library can meter and price. */
+declare const AI_OPERATIONS: readonly ["chat", "responses", "embeddings", "image", "video", "audio", "rerank", "moderation"];
+/** A logical AI operation kind. */
+type AiOperation = (typeof AI_OPERATIONS)[number];
+
+/**
+ * @fileoverview Catalog of response service tiers. Providers may silently
+ * downgrade a request (OpenAI returns `service_tier: "default"` when priority
+ * capacity is exhausted), so pricing keys off the tier reported in the RESPONSE,
+ * which is what the normalizers read (see spec §5.1).
+ * @layer shared
+ */
+/**
+ * Service tier of the actual response. `'batch'` is the Batch API (~50% discount);
+ * `'flex'` is OpenAI flex processing (batch rates, synchronous); `'priority'` is
+ * paid premium; `'standard'` is the default.
+ */
+declare const SERVICE_TIERS: readonly ["standard", "batch", "flex", "priority"];
+/** A response service tier. */
+type ServiceTier = (typeof SERVICE_TIERS)[number];
+
+/**
+ * @fileoverview Catalog of the AI providers this library recognizes out of the
+ * box. Custom OpenAI-compatible providers register their own id at runtime via
+ * `providerPresets.openaiCompatible(id)`, so the public {@link ProviderId} type
+ * (see `types/catalogs.ts`) widens this closed list with `(string & {})`.
+ * @layer shared
+ */
+/**
+ * The provider identifiers shipped with the library. Frozen tuple so the derived
+ * {@link KnownProviderId} union stays in sync with the runtime values.
+ */
+declare const PROVIDER_IDS: readonly ["openai", "azure-openai", "anthropic", "gemini", "vertex", "mistral", "bedrock", "openrouter", "deepseek", "xai", "groq"];
+/** A provider id known to the library at build time (excludes custom ids). */
+type KnownProviderId = (typeof PROVIDER_IDS)[number];
+
+/**
+ * @fileoverview Canonical catalog types that the rest of the library is generic
+ * over: provider/operation/tier unions, the metering scope, the rating mode, and
+ * the pure normalizer/preset contracts. Framework-free.
+ * @layer shared
+ */
+
+/**
+ * A provider identifier. Widened with `(string & {})` so custom
+ * OpenAI-compatible providers can register their own id while known ids keep
+ * editor autocomplete.
+ */
+type ProviderId = KnownProviderId | (string & {});
+
+/**
+ * @fileoverview The effective-dated price row and its insert type. Rates are
+ * integer nano-USD per 1,000,000 tokens (or per unit for non-token line items),
+ * so every cost computation is exact integer math.
+ * @layer shared
+ */
+
+/**
+ * One immutable, effective-dated price version for a (provider, model, operation,
+ * serviceTier) tuple. A call is rated at the row in effect at its timestamp.
+ */
+interface PriceVersion {
+    id: string;
+    provider: ProviderId;
+    model: string;
+    operation: AiOperation;
+    /** Part of the resolution key; `'standard'` is the default. */
+    serviceTier: ServiceTier;
+    /** Rates as integer nano-USD per 1,000,000 tokens. */
+    inputNanoUsdPerMillion: bigint;
+    outputNanoUsdPerMillion: bigint;
+    cacheReadNanoUsdPerMillion: bigint;
+    cacheWrite5mNanoUsdPerMillion: bigint;
+    cacheWrite1hNanoUsdPerMillion: bigint;
+    /** Usually equals output; kept separate for models that price reasoning differently. */
+    reasoningNanoUsdPerMillion: bigint;
+    audioInNanoUsdPerMillion: bigint;
+    audioOutNanoUsdPerMillion: bigint;
+    imageInNanoUsdPerMillion: bigint;
+    imageOutNanoUsdPerMillion: bigint;
+    /**
+     * Long-context tier threshold: when total input (all input-side categories)
+     * exceeds this, the tier rates replace the base input/output rates for the
+     * WHOLE call (all-or-nothing, matching Gemini's billing).
+     */
+    tierThresholdTokens?: number;
+    tierInputNanoUsdPerMillion?: bigint;
+    tierOutputNanoUsdPerMillion?: bigint;
+    /**
+     * Non-token line items in nano-USD PER UNIT, matched against
+     * `NormalizedUsage.serverToolUse` / `MeteringContext.extraUnits` by EXACT key —
+     * so the key must equal the count key the normalizer emits (e.g. Anthropic's
+     * `usage.server_tool_use.web_search_requests`) — e.g.
+     * `{ web_search_requests: 10_000_000n }` ($0.01 per web-search call).
+     */
+    unitRates?: Record<string, bigint>;
+    currency: 'USD';
+    effectiveFrom: Date;
+    /** `null` = current; set when a newer row supersedes it. */
+    effectiveTo: Date | null;
+    /** Provenance: `'snapshot' | 'manual' | 'import'`. */
+    source: string;
+}
+
+/**
+ * @fileoverview Pinned per-model price snapshot in bigint nano-USD per 1,000,000
+ * tokens (unit rates are per unit). Converted offline from published provider
+ * pricing (methodology mirrors LiteLLM's `model_prices_and_context_window.json`,
+ * see `scripts/convert-litellm-prices.mjs`). Snapshot date: 2026-07. Pinning
+ * keeps rates point-in-time stable; hosts add or override models via
+ * `PricingService.upsertPrice()`. Data only — the exported array carries no
+ * behavior. Each row's `source` is `'snapshot'`; rates flagged `VERIFY` should be
+ * spot-checked against the provider's pricing page during review.
+ * @layer prices
+ */
+
+/** A seed price row: a {@link PriceVersion} without the store-assigned identity/effective-date fields. */
+type SeedPriceRow = Omit<PriceVersion, 'id' | 'effectiveFrom' | 'effectiveTo'>;
+/**
+ * The pinned price snapshot. Rated at published standard-tier prices unless the
+ * row's `serviceTier` says otherwise; batch rows are ~50% of standard.
+ */
+declare const MODEL_PRICES_SEED: readonly SeedPriceRow[];
+
+export { MODEL_PRICES_SEED, type SeedPriceRow };