26cb0cc2e1
Expose Kimi K3 throughout the native pi catalog and token-cost surfaces. - Add K3 pricing and normalized lookup coverage for Dashboard cost calculations. - Verify native K3 registry, API model route, and repeat-stop behavior end to end. - Add a minor release changeset for Kimi K3 model selection support. Files changed: .changeset/fn-8180-pi-sdk-kimi-k3.md | 7 +++ packages/core/src/__tests__/model-pricing.test.ts | 7 ++- packages/core/src/model-pricing.ts | 17 +++++- ...ister-model-routes-kimi-k3-supplemental.test.ts | 70 ++++++++++++++++++++++ .../pi-prompt-session-and-check-recursion.test.ts | 15 +++++ .../src/__tests__/provider-registration.test.ts | 17 ++++++ 6 files changed, 130 insertions(+), 3 deletions(-) Fusion-Task-Id: FN-8180 Fusion-Task-Lineage: 07598d73-c866-4329-b5cf-3aebc1c0a180 Co-authored-by: Fusion (runfusion.ai) <noreply@runfusion.ai>
571 lines
21 KiB
TypeScript
571 lines
21 KiB
TypeScript
/**
|
||
* Model pricing → USD cost derivation (KTD6, U3).
|
||
*
|
||
* Cost is **derived at read time** from token counts × a hand-maintained
|
||
* pricing map plus optional user-managed overrides; it is never persisted (so
|
||
* historical rows stay correct when prices change, and no backfill migration is
|
||
* needed). Unknown models surface tokens with cost marked `unavailable` rather
|
||
* than guessing a price.
|
||
*
|
||
* ⚠️ HAND-MAINTAINED MAP. The `MODEL_PRICING` table below is curated by humans
|
||
* from each provider's public pricing pages — it is NOT fetched at runtime.
|
||
* Callers may supply persisted overrides, including entries parsed from the
|
||
* canonical LiteLLM dataset, and those overrides take precedence over this
|
||
* baseline. When you update a baseline rate, bump {@link pricingAsOf} in the
|
||
* same change. The UI surfaces `pricingAsOf` ("prices as of <date>") and marks
|
||
* entries older than {@link PRICING_STALE_AFTER_MS} as low-confidence, so
|
||
* stale-but-present rates (which the unknown-model guard does not catch) are
|
||
* visible rather than silently wrong.
|
||
*
|
||
* Rates are USD **per 1,000,000 tokens**.
|
||
*
|
||
* Pure data module: no DB, no I/O, and no `Date.now()` at import time. Callers
|
||
* that care about staleness pass an explicit `now`; otherwise staleness is
|
||
* judged against {@link pricingAsOf} alone (i.e. never stale).
|
||
*/
|
||
|
||
/**
|
||
* The date the rates in {@link MODEL_PRICING} were last verified, ISO-8601.
|
||
* Bump this whenever you edit a rate. Surfaced in the UI as "prices as of".
|
||
*/
|
||
export const pricingAsOf = "2026-07-16";
|
||
|
||
/**
|
||
* Pricing entries older than this (relative to a caller-supplied `now`) are
|
||
* flagged `stale: true`. 180 days ≈ two quarters — long enough that routine
|
||
* price churn doesn't fire constantly, short enough that a long-unmaintained
|
||
* map is surfaced. Compared against {@link pricingAsOf}, not per-entry dates.
|
||
*/
|
||
export const PRICING_STALE_AFTER_MS = 180 * 24 * 60 * 60 * 1000;
|
||
|
||
/*
|
||
* FNXC:CommandCenter 2026-06-22-00:00:
|
||
* Users need one-click pricing refreshes from LiteLLM's continuously updated community dataset while the core module remains pure. Keep the URL as data only; dashboard routes own HTTP, validation errors, and persistence.
|
||
*/
|
||
export const LITELLM_PRICING_SOURCE_URL =
|
||
"https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json";
|
||
|
||
export const LITELLM_PRICING_SOURCE_LABEL = "litellm/model_prices_and_context_window.json";
|
||
|
||
/** A single model's per-1M-token rates plus a citation. */
|
||
export interface ModelPricing {
|
||
/** USD per 1M uncached input tokens. */
|
||
inputPer1M: number;
|
||
/** USD per 1M output tokens. */
|
||
outputPer1M: number;
|
||
/** USD per 1M cache-read (cached) input tokens. */
|
||
cacheReadPer1M: number;
|
||
/** USD per 1M cache-write tokens. */
|
||
cacheWritePer1M: number;
|
||
/** Where the rate came from (provider pricing page / docs). */
|
||
source: string;
|
||
}
|
||
|
||
/** User-managed pricing overrides keyed by lowercased `provider:model`. */
|
||
export type ModelPricingOverrides = Record<string, ModelPricing>;
|
||
|
||
/** Token counts to price. Mirrors {@link TokenTotals} from token-analytics. */
|
||
export interface UsageForCost {
|
||
inputTokens: number;
|
||
outputTokens: number;
|
||
/** Cache-read tokens (priced at the cache-read rate, NOT the input rate). */
|
||
cachedTokens: number;
|
||
/** Cache-write tokens (priced at the cache-write rate). */
|
||
cacheWriteTokens: number;
|
||
}
|
||
|
||
/** Result of {@link costFor}. */
|
||
export interface CostResult {
|
||
/** Derived USD cost, or `null` when no price is known for the model. */
|
||
usd: number | null;
|
||
/** True when the model has no pricing entry (cost is a guess-free `null`). */
|
||
unavailable: boolean;
|
||
/** True when the pricing map is older than the staleness threshold. */
|
||
stale: boolean;
|
||
}
|
||
|
||
/**
|
||
* Hand-maintained pricing table, keyed by `provider:model`.
|
||
*
|
||
* Keys are lowercased `${provider}:${model}`. Lookup also falls back to the
|
||
* bare model id (`:model`) so callers that only know the model still resolve.
|
||
* Model ids match the strings Fusion stores in `tasks.modelId` /
|
||
* `tasks.modelProvider` (see `runtime-provider-probes.ts` and grep for
|
||
* `modelId`/`modelProvider`): Anthropic Claude, OpenAI, Google Gemini.
|
||
*
|
||
* Sources (verified 2026-06-15, see `pricingAsOf`):
|
||
* - Anthropic: platform.claude.com/docs/en/pricing (per-MTok; cache read ≈
|
||
* 0.1× input, 5-min cache write ≈ 1.25× input).
|
||
* - OpenAI: openai.com/api/pricing (cached input ≈ 0.5×/0.25× input; OpenAI
|
||
* has no separate cache-write charge, so cacheWrite = input rate).
|
||
* - Google Gemini: ai.google.dev/gemini-api/docs/pricing (context-cache read
|
||
* rate; no distinct cache-write token charge, so cacheWrite = input rate).
|
||
*/
|
||
export const MODEL_PRICING: Readonly<Record<string, ModelPricing>> = {
|
||
// ── Anthropic Claude ────────────────────────────────────────────────
|
||
// input / output / cacheRead(0.1×) / cacheWrite(1.25×, 5-min TTL)
|
||
/*
|
||
* FNXC:ModelCatalog 2026-07-01-22:40:
|
||
* `anthropic:claude-sonnet-5` is advertised again (works on raw API key + Claude CLI; live-verified), so restore its static pricing. Matches the cost in SUPPLEMENTAL_ANTHROPIC_PROVIDER_REGISTRATION.
|
||
*/
|
||
"anthropic:claude-sonnet-5": {
|
||
inputPer1M: 2,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.2,
|
||
cacheWritePer1M: 2.5,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-8": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 25,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 6.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-7": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 25,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 6.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-6": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 25,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 6.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-5": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 25,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 6.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-1": {
|
||
inputPer1M: 15,
|
||
outputPer1M: 75,
|
||
cacheReadPer1M: 1.5,
|
||
cacheWritePer1M: 18.75,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-opus-4-20250514": {
|
||
inputPer1M: 15,
|
||
outputPer1M: 75,
|
||
cacheReadPer1M: 1.5,
|
||
cacheWritePer1M: 18.75,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-sonnet-4-6": {
|
||
inputPer1M: 3,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.3,
|
||
cacheWritePer1M: 3.75,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-sonnet-4-5": {
|
||
inputPer1M: 3,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.3,
|
||
cacheWritePer1M: 3.75,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-sonnet-4-20250514": {
|
||
inputPer1M: 3,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.3,
|
||
cacheWritePer1M: 3.75,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-haiku-4-5": {
|
||
inputPer1M: 1,
|
||
outputPer1M: 5,
|
||
cacheReadPer1M: 0.1,
|
||
cacheWritePer1M: 1.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-haiku-4-5-20251001": {
|
||
inputPer1M: 1,
|
||
outputPer1M: 5,
|
||
cacheReadPer1M: 0.1,
|
||
cacheWritePer1M: 1.25,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
"anthropic:claude-fable-5": {
|
||
inputPer1M: 10,
|
||
outputPer1M: 50,
|
||
cacheReadPer1M: 1,
|
||
cacheWritePer1M: 12.5,
|
||
source: "platform.claude.com/docs/en/pricing",
|
||
},
|
||
|
||
// ── OpenAI ──────────────────────────────────────────────────────────
|
||
// OpenAI has no separate cache-write charge → cacheWrite = input rate.
|
||
"openai:gpt-5": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.125,
|
||
cacheWritePer1M: 1.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:gpt-5-mini": {
|
||
inputPer1M: 0.25,
|
||
outputPer1M: 2,
|
||
cacheReadPer1M: 0.025,
|
||
cacheWritePer1M: 0.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:gpt-4o": {
|
||
inputPer1M: 2.5,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 1.25,
|
||
cacheWritePer1M: 2.5,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:gpt-4o-mini": {
|
||
inputPer1M: 0.15,
|
||
outputPer1M: 0.6,
|
||
cacheReadPer1M: 0.075,
|
||
cacheWritePer1M: 0.15,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:gpt-4.1": {
|
||
inputPer1M: 2,
|
||
outputPer1M: 8,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 2,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:gpt-4-turbo": {
|
||
inputPer1M: 10,
|
||
outputPer1M: 30,
|
||
cacheReadPer1M: 10,
|
||
cacheWritePer1M: 10,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:o1": {
|
||
inputPer1M: 15,
|
||
outputPer1M: 60,
|
||
cacheReadPer1M: 7.5,
|
||
cacheWritePer1M: 15,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai:o3-mini": {
|
||
inputPer1M: 1.1,
|
||
outputPer1M: 4.4,
|
||
cacheReadPer1M: 0.55,
|
||
cacheWritePer1M: 1.1,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
|
||
// ── OpenAI Codex ────────────────────────────────────────────────────
|
||
// OpenAI has no separate cache-write charge → cacheWrite = input rate.
|
||
/*
|
||
* FNXC:CommandCenter 2026-06-21-12:14:
|
||
* Codex runs store the `openai-codex` provider, so pricing must be keyed as `openai-codex:<modelId>` instead of relying on the OpenAI provider or bare-model fallback. Keep these entries explicit so Command Center token cost does not show `unavailable`; rates mirror OpenAI GPT-5 Codex pricing, and `pricingAsOf` must be bumped on every rate edit.
|
||
*/
|
||
"openai-codex:gpt-5-codex": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.125,
|
||
cacheWritePer1M: 1.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.1-codex": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.125,
|
||
cacheWritePer1M: 1.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.2-codex": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.125,
|
||
cacheWritePer1M: 1.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.3-codex": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.125,
|
||
cacheWritePer1M: 1.25,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
/*
|
||
* FNXC:CommandCenter 2026-07-09-21:30:
|
||
* FN-7757 found that pi-ai 0.80.5 now persists OpenAI Codex model ids from the generated `openai-codex.models.js` catalog (`gpt-5.3-codex-spark`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.5`, and codenamed GPT-5.6 variants). The old static table covered only earlier `gpt-5.x-codex` ids, so lookupPricing returned unavailable and every dashboard cost surface rendered `—`. Keep this block aligned with the pinned runtime catalog rates while preserving unknown-model `—` semantics.
|
||
*/
|
||
"openai-codex:gpt-5.3-codex-spark": {
|
||
inputPer1M: 1.75,
|
||
outputPer1M: 14,
|
||
cacheReadPer1M: 0.175,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.4": {
|
||
inputPer1M: 2.5,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.25,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.4-mini": {
|
||
inputPer1M: 0.75,
|
||
outputPer1M: 4.5,
|
||
cacheReadPer1M: 0.075,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.5": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 30,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.6-luna": {
|
||
inputPer1M: 1,
|
||
outputPer1M: 6,
|
||
cacheReadPer1M: 0.1,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.6-sol": {
|
||
inputPer1M: 5,
|
||
outputPer1M: 30,
|
||
cacheReadPer1M: 0.5,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:gpt-5.6-terra": {
|
||
inputPer1M: 2.5,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.25,
|
||
cacheWritePer1M: 0,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
"openai-codex:codex-mini-latest": {
|
||
inputPer1M: 1.5,
|
||
outputPer1M: 6,
|
||
cacheReadPer1M: 0.375,
|
||
cacheWritePer1M: 1.5,
|
||
source: "openai.com/api/pricing",
|
||
},
|
||
|
||
// ── Google Gemini ───────────────────────────────────────────────────
|
||
// No distinct cache-write token charge → cacheWrite = input rate.
|
||
"google:gemini-2.5-pro": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.31,
|
||
cacheWritePer1M: 1.25,
|
||
source: "ai.google.dev/gemini-api/docs/pricing",
|
||
},
|
||
"google:gemini-2.5-flash": {
|
||
inputPer1M: 0.3,
|
||
outputPer1M: 2.5,
|
||
cacheReadPer1M: 0.075,
|
||
cacheWritePer1M: 0.3,
|
||
source: "ai.google.dev/gemini-api/docs/pricing",
|
||
},
|
||
"google:gemini-2.0-flash": {
|
||
inputPer1M: 0.1,
|
||
outputPer1M: 0.4,
|
||
cacheReadPer1M: 0.025,
|
||
cacheWritePer1M: 0.1,
|
||
source: "ai.google.dev/gemini-api/docs/pricing",
|
||
},
|
||
"google:gemini-2.0-pro": {
|
||
inputPer1M: 1.25,
|
||
outputPer1M: 10,
|
||
cacheReadPer1M: 0.31,
|
||
cacheWritePer1M: 1.25,
|
||
source: "ai.google.dev/gemini-api/docs/pricing",
|
||
},
|
||
|
||
/*
|
||
* FNXC:ModelCatalog 2026-07-11-23:02:
|
||
* The LiteLLM pricing refresh only maps OpenAI, Anthropic, and Google providers, so Z.ai, MiniMax, and Kimi Coding runs need static rows to keep Dashboard token-cost surfaces from rendering `—`. Rates are verified from https://docs.z.ai/guides/overview/pricing.md, https://platform.minimax.io/docs/guides/pricing-paygo.md, https://platform.kimi.ai/docs/pricing/chat-k26.md, and https://platform.kimi.ai/docs/pricing/chat-k3.md; MiniMax uses the Standard ≤512k pay-as-you-go tier because MODEL_PRICING has no request-tier dimension.
|
||
|
||
* FNXC:ModelCatalog 2026-07-16-19:05:
|
||
* FN-8180 upgrades pi to 0.80.10, whose native kimi-coding catalog exposes `k3`.
|
||
* Keep the Dashboard's independent cost derivation aligned with Moonshot's published
|
||
* K3 rates rather than pi's intentionally zero-valued catalog cost placeholders.
|
||
*/
|
||
// ── Zhipu AI (Z.ai) ─────────────────────────────────────────────────
|
||
// No distinct cache-write token charge → cacheWrite = input rate.
|
||
"zai:glm-5.2": {
|
||
inputPer1M: 1.4,
|
||
outputPer1M: 4.4,
|
||
cacheReadPer1M: 0.26,
|
||
cacheWritePer1M: 1.4,
|
||
source: "docs.z.ai/guides/overview/pricing.md",
|
||
},
|
||
|
||
// ── MiniMax ─────────────────────────────────────────────────────────
|
||
// Standard pay-as-you-go ≤512k tier; no M3 cache-write row → cacheWrite = input rate.
|
||
"minimax:minimax-m3": {
|
||
inputPer1M: 0.3,
|
||
outputPer1M: 1.2,
|
||
cacheReadPer1M: 0.06,
|
||
cacheWritePer1M: 0.3,
|
||
source: "platform.minimax.io/docs/guides/pricing-paygo.md",
|
||
},
|
||
|
||
// ── Moonshot AI (Kimi) ──────────────────────────────────────────────
|
||
// Cache miss is the input rate; no distinct cache-write charge → cacheWrite = input rate.
|
||
"kimi-coding:kimi-k2.6-preview": {
|
||
inputPer1M: 0.95,
|
||
outputPer1M: 4,
|
||
cacheReadPer1M: 0.16,
|
||
cacheWritePer1M: 0.95,
|
||
source: "platform.kimi.ai/docs/pricing/chat-k26.md",
|
||
},
|
||
// Cache miss is the input rate; K3 has no separately published cache-write rate.
|
||
"kimi-coding:k3": {
|
||
inputPer1M: 3,
|
||
outputPer1M: 15,
|
||
cacheReadPer1M: 0.3,
|
||
cacheWritePer1M: 3,
|
||
source: "platform.kimi.ai/docs/pricing/chat-k3.md",
|
||
},
|
||
};
|
||
|
||
/** Reference to a model, by provider + id (either may be unset). */
|
||
export interface ModelRef {
|
||
provider?: string | null;
|
||
model?: string | null;
|
||
}
|
||
|
||
function normalize(s: string | null | undefined): string {
|
||
return (s ?? "").trim().toLowerCase();
|
||
}
|
||
|
||
function findBareModelPricing(
|
||
model: string,
|
||
entries: Record<string, ModelPricing> | Readonly<Record<string, ModelPricing>>,
|
||
): ModelPricing | undefined {
|
||
for (const [key, entry] of Object.entries(entries)) {
|
||
if (key.endsWith(`:${model}`)) return entry;
|
||
}
|
||
return undefined;
|
||
}
|
||
|
||
/**
|
||
* Resolve a pricing entry for a model. Tries override `provider:model` first,
|
||
* then override bare-model fallback, then the built-in baseline using the same
|
||
* precedence. Returns `undefined` for unknown models — callers must treat that
|
||
* as `unavailable`, never as a guessed price.
|
||
*
|
||
* FNXC:CommandCenter 2026-06-22-00:00:
|
||
* Editable/fetched model rates must override the hand-maintained baseline without removing the baseline fallback. Keep exact provider:model checks before bare-model scans so provider-specific overrides stay deterministic.
|
||
*/
|
||
export function lookupPricing(ref: ModelRef, overrides?: ModelPricingOverrides): ModelPricing | undefined {
|
||
const provider = normalize(ref.provider);
|
||
const model = normalize(ref.model);
|
||
if (!model) return undefined;
|
||
if (provider) {
|
||
const exactOverride = overrides?.[`${provider}:${model}`];
|
||
if (exactOverride) return exactOverride;
|
||
}
|
||
const bareOverride = overrides ? findBareModelPricing(model, overrides) : undefined;
|
||
if (bareOverride) return bareOverride;
|
||
if (provider) {
|
||
const exact = MODEL_PRICING[`${provider}:${model}`];
|
||
if (exact) return exact;
|
||
}
|
||
return findBareModelPricing(model, MODEL_PRICING);
|
||
}
|
||
|
||
function litellmProviderToFusionProvider(provider: unknown): string | null {
|
||
if (typeof provider !== "string") return null;
|
||
const normalized = normalize(provider);
|
||
if (normalized === "openai") return "openai";
|
||
if (normalized === "anthropic") return "anthropic";
|
||
if (normalized === "gemini" || normalized.startsWith("gemini")) return "google";
|
||
if (normalized === "vertex_ai" || normalized.startsWith("vertex_ai")) return "google";
|
||
if (normalized === "vertex_ai-language-models") return "google";
|
||
return null;
|
||
}
|
||
|
||
function numericField(entry: Record<string, unknown>, key: string): number | undefined {
|
||
const value = entry[key];
|
||
return typeof value === "number" && Number.isFinite(value) ? value : undefined;
|
||
}
|
||
|
||
/**
|
||
* Parse LiteLLM's canonical pricing dataset into Fusion pricing overrides.
|
||
* Pure: no HTTP, DB access, or clock reads. Unsupported providers and non-chat
|
||
* rows are skipped so a broad upstream dataset can safely feed Fusion's known
|
||
* model-provider surface.
|
||
*/
|
||
export function parseLiteLLMPricing(json: unknown): { overrides: ModelPricingOverrides; count: number } {
|
||
const overrides: ModelPricingOverrides = {};
|
||
if (json === null || typeof json !== "object" || Array.isArray(json)) {
|
||
return { overrides, count: 0 };
|
||
}
|
||
for (const [modelId, value] of Object.entries(json as Record<string, unknown>)) {
|
||
if (modelId === "sample_spec") continue;
|
||
if (value === null || typeof value !== "object" || Array.isArray(value)) continue;
|
||
const entry = value as Record<string, unknown>;
|
||
if (entry.mode !== "chat") continue;
|
||
const provider = litellmProviderToFusionProvider(entry.litellm_provider);
|
||
if (!provider) continue;
|
||
const inputCost = numericField(entry, "input_cost_per_token");
|
||
const outputCost = numericField(entry, "output_cost_per_token");
|
||
if (inputCost === undefined || outputCost === undefined) continue;
|
||
const inputPer1M = inputCost * 1_000_000;
|
||
const outputPer1M = outputCost * 1_000_000;
|
||
const cacheRead = numericField(entry, "cache_read_input_token_cost");
|
||
const cacheWrite = numericField(entry, "cache_creation_input_token_cost");
|
||
overrides[`${provider}:${normalize(modelId)}`] = {
|
||
inputPer1M,
|
||
outputPer1M,
|
||
cacheReadPer1M: cacheRead === undefined ? inputPer1M : cacheRead * 1_000_000,
|
||
cacheWritePer1M: cacheWrite === undefined ? inputPer1M : cacheWrite * 1_000_000,
|
||
source: LITELLM_PRICING_SOURCE_LABEL,
|
||
};
|
||
}
|
||
return { overrides, count: Object.keys(overrides).length };
|
||
}
|
||
|
||
/** True when the pricing map is older than the threshold relative to `now`. */
|
||
function isStale(now: number | undefined): boolean {
|
||
if (now === undefined) return false;
|
||
const asOf = Date.parse(pricingAsOf);
|
||
if (Number.isNaN(asOf)) return false;
|
||
return now - asOf > PRICING_STALE_AFTER_MS;
|
||
}
|
||
|
||
/**
|
||
* Derive USD cost for `usage` under `model`'s rates.
|
||
*
|
||
* - Unknown model → `{ usd: null, unavailable: true, stale }` (never guessed).
|
||
* - Cache-read tokens are priced at the cache-read rate, cache-write tokens at
|
||
* the cache-write rate — NOT the input rate.
|
||
* - `stale` is true when the (caller-supplied) `now` is more than
|
||
* {@link PRICING_STALE_AFTER_MS} past {@link pricingAsOf}. With no `now`,
|
||
* `stale` is always false.
|
||
*/
|
||
export function costFor(
|
||
usage: UsageForCost,
|
||
model: ModelRef,
|
||
now?: number,
|
||
overrides?: ModelPricingOverrides,
|
||
): CostResult {
|
||
const stale = isStale(now);
|
||
const pricing = lookupPricing(model, overrides);
|
||
if (!pricing) {
|
||
return { usd: null, unavailable: true, stale };
|
||
}
|
||
const usd =
|
||
(usage.inputTokens * pricing.inputPer1M +
|
||
usage.outputTokens * pricing.outputPer1M +
|
||
usage.cachedTokens * pricing.cacheReadPer1M +
|
||
usage.cacheWriteTokens * pricing.cacheWritePer1M) /
|
||
1_000_000;
|
||
return { usd, unavailable: false, stale };
|
||
}
|