feat(ai-openrouter): surface per-request cost on RUN_FINISHED

.changeset/openrouter-cost-tracking.md

@@ -0,0 +1,22 @@
+---
+'@tanstack/ai-openrouter': minor
+'@tanstack/ai': patch
+---
+
+Surface OpenRouter's per-request cost on `RUN_FINISHED.usage`.
+
+OpenRouter reports the actual cost of each request inline on the chat response.
+The `openRouterText` and `openRouterResponsesText` adapters now forward that
+value on the terminal `RUN_FINISHED` event as `usage.cost`, with OpenRouter's
+per-request breakdown under `usage.costDetails`. This is the cost OpenRouter
+itself reports — it is not computed locally from token counts, so it accounts
+for routing, fallback providers, BYOK, and cached-token pricing.
+
+`@tanstack/ai` adds a shared `UsageTotals` type with optional `cost` and
+`costDetails` fields, plus a `UsageCostDetails` type enumerating the known
+breakdown fields (`upstreamInferenceCost`, `upstreamInferencePromptCost`,
+`upstreamInferenceCompletionsCost`, `upstreamInferenceInputCost`,
+`upstreamInferenceOutputCost`). `RunFinishedEvent.usage`, the middleware
+`UsageInfo` (`onUsage`), and `FinishInfo.usage` (`onFinish`) all use it, so cost
+can be read without casts and with autocomplete on the breakdown. The fields
+are optional and additive — adapters that do not report cost are unaffected.

README.md

@@ -186,7 +186,7 @@ Official adapters include:
 
 | Package                                                                              | Use it for                                                                     |
 | ------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------ |
-| [`@tanstack/ai-openrouter`](https://tanstack.com/ai/latest/docs/adapters/openrouter) | 300+ models through one OpenRouter API                                         |
+| [`@tanstack/ai-openrouter`](https://tanstack.com/ai/latest/docs/adapters/openrouter) | 300+ models through one OpenRouter API, with per-request cost tracking         |
 | [`@tanstack/ai-openai`](https://tanstack.com/ai/latest/docs/adapters/openai)         | OpenAI chat, image, video, speech, transcription, realtime, and provider tools |
 | [`@tanstack/ai-anthropic`](https://tanstack.com/ai/latest/docs/adapters/anthropic)   | Anthropic Claude chat, thinking, tools, and structured outputs                 |
 | [`@tanstack/ai-gemini`](https://tanstack.com/ai/latest/docs/adapters/gemini)         | Google Gemini chat, image, speech, and audio generation                        |

docs/adapters/openrouter.md

@@ -169,6 +169,38 @@ Caveats while the Responses adapter is in beta:
 - If in doubt, prefer `openRouterText`. The Chat Completions endpoint has
   broader provider coverage and feature parity today.
 
+## Cost Tracking
+
+OpenRouter reports the actual cost of each request inline on the streamed
+response. When present, the adapter forwards it on the terminal `RUN_FINISHED`
+event under `usage.cost`, with OpenRouter's per-request breakdown under
+`usage.costDetails`. This is the cost OpenRouter itself reports for the
+request — it is **not** computed locally from token counts, so it already
+accounts for routing, fallback providers, BYOK, and cached-token pricing. See
+OpenRouter's [Usage Accounting](https://openrouter.ai/docs/use-cases/usage-accounting)
+docs for the meaning and units of these fields.
+
+```typescript
+import { chat } from "@tanstack/ai";
+import { openRouterText } from "@tanstack/ai-openrouter";
+
+for await (const chunk of chat({
+  adapter: openRouterText("openai/gpt-5"),
+  messages: [{ role: "user", content: "Hello!" }],
+})) {
+  if (chunk.type === "RUN_FINISHED") {
+    console.log("cost:", chunk.usage?.cost);
+    console.log("breakdown:", chunk.usage?.costDetails);
+  }
+}
+```
+
+The same `usage` (including `cost` / `costDetails`) is passed to middleware via
+the `onUsage` and `onFinish` hooks. When OpenRouter does not report a cost, the
+fields are simply absent and the stream completes normally. Both
+`openRouterText` and `openRouterResponsesText` populate cost when OpenRouter
+returns it.
+
 ## Next Steps
 
 - [Getting Started](../getting-started/quick-start) - Learn the basics

packages/ai-openrouter/src/adapters/cost.ts

@@ -0,0 +1,92 @@
+/**
+ * Helpers for extracting OpenRouter's provider-reported per-request cost from the
+ * SDK usage object and shaping it for `RUN_FINISHED.usage`.
+ *
+ * OpenRouter returns an authoritative per-request `cost` plus an optional
+ * `cost_details` breakdown. We forward the known breakdown fields verbatim — no
+ * local token-times-price math, which would drift across routing/fallback/BYOK
+ * /caching. Unknown breakdown keys are dropped so the public `UsageCostDetails`
+ * shape stays closed.
+ *
+ * Input is intentionally typed `unknown`: callers pass usage objects whose static
+ * types are narrowed to token-only fields (notably the Responses adapter), and the
+ * Responses usage normalizer can leave `cost_details` in snake_case. Reading both
+ * `costDetails` and `cost_details` and narrowing here keeps every call site simple.
+ */
+
+import type { UsageCostDetails } from '@tanstack/ai'
+
+export interface ExtractedCost {
+  cost?: number
+  costDetails?: UsageCostDetails
+}
+
+/** Keys recognized in `cost_details` / `costDetails`, snake_case ↔ camelCase. */
+const KNOWN_DETAIL_KEYS: Record<string, keyof UsageCostDetails> = {
+  upstream_inference_cost: 'upstreamInferenceCost',
+  upstreamInferenceCost: 'upstreamInferenceCost',
+  upstream_inference_prompt_cost: 'upstreamInferencePromptCost',
+  upstreamInferencePromptCost: 'upstreamInferencePromptCost',
+  upstream_inference_completions_cost: 'upstreamInferenceCompletionsCost',
+  upstreamInferenceCompletionsCost: 'upstreamInferenceCompletionsCost',
+  upstream_inference_input_cost: 'upstreamInferenceInputCost',
+  upstreamInferenceInputCost: 'upstreamInferenceInputCost',
+  upstream_inference_output_cost: 'upstreamInferenceOutputCost',
+  upstreamInferenceOutputCost: 'upstreamInferenceOutputCost',
+}
+
+function asRecord(value: unknown): Record<string, unknown> | undefined {
+  return typeof value === 'object' && value !== null
+    ? (value as Record<string, unknown>)
+    : undefined
+}
+
+/**
+ * Narrow a raw `cost_details`/`costDetails` map to the known fields of
+ * `UsageCostDetails`. Negative values (e.g. discounts) are preserved; `null`,
+ * non-finite numbers, non-numeric values, and unknown keys are dropped.
+ */
+function extractCostDetails(details: unknown): UsageCostDetails | undefined {
+  const record = asRecord(details)
+  if (!record) return undefined
+
+  const out: UsageCostDetails = {}
+  for (const [rawKey, value] of Object.entries(record)) {
+    const key = KNOWN_DETAIL_KEYS[rawKey]
+    if (!key) continue
+    if (typeof value === 'number' && Number.isFinite(value)) {
+      out[key] = value
+    }
+  }
+
+  return Object.keys(out).length > 0 ? out : undefined
+}
+
+/**
+ * Extract `cost`/`costDetails` from a provider usage object.
+ *
+ * - `cost` is attached only when it is a finite number — this preserves `cost === 0`
+ *   and rejects `NaN`/`Infinity`, and does not clamp negative values.
+ * - `costDetails` is attached only alongside a valid `cost` (an orphan breakdown
+ *   without a total cannot be reconciled and is dropped). Both camelCase
+ *   `costDetails` and snake_case `cost_details` are read.
+ *
+ * Returns an empty object when no usable cost is present, so call sites can spread
+ * the result unconditionally.
+ */
+export function extractUsageCost(usage: unknown): ExtractedCost {
+  const record = asRecord(usage)
+  if (!record) return {}
+
+  const cost = record.cost
+  if (typeof cost !== 'number' || !Number.isFinite(cost)) return {}
+
+  const costDetails = extractCostDetails(
+    record.costDetails ?? record.cost_details,
+  )
+
+  return {
+    cost,
+    ...(costDetails && { costDetails }),
+  }
+}

packages/ai-openrouter/src/adapters/responses-text.ts

@@ -9,6 +9,7 @@ import { convertFunctionToolToResponsesFormat } from '../internal/responses-tool
 import { isWebSearchTool } from '../tools/web-search-tool'
 import { isWebFetchTool } from '../tools/web-fetch-tool'
 import { getOpenRouterApiKeyFromEnv } from '../utils'
+import { extractUsageCost } from './cost'
 import type { SDKOptions } from '@openrouter/sdk'
 import type { ResponsesFunctionTool } from '../internal/responses-tool-converter'
 import type {
@@ -623,6 +624,7 @@ export class OpenRouterResponsesTextAdapter<
             promptTokens: usage.inputTokens ?? 0,
             completionTokens: usage.outputTokens ?? 0,
             totalTokens: usage.totalTokens ?? 0,
+            ...extractUsageCost(usage),
           },
         }),
       }
@@ -1433,6 +1435,7 @@ export class OpenRouterResponsesTextAdapter<
               promptTokens: responseObj.usage?.inputTokens || 0,
               completionTokens: responseObj.usage?.outputTokens || 0,
               totalTokens: responseObj.usage?.totalTokens || 0,
+              ...extractUsageCost(responseObj.usage),
             },
             finishReason,
           }

packages/ai-openrouter/src/adapters/text.ts

@@ -7,6 +7,7 @@ import { extractRequestOptions } from '../internal/request-options'
 import { makeStructuredOutputCompatible } from '../internal/schema-converter'
 import { convertToolsToProviderFormat } from '../tools'
 import { getOpenRouterApiKeyFromEnv } from '../utils'
+import { extractUsageCost } from './cost'
 import type { SDKOptions } from '@openrouter/sdk'
 import type {
   ChatContentItems,
@@ -549,6 +550,7 @@ export class OpenRouterTextAdapter<
             promptTokens: lastUsage.promptTokens,
             completionTokens: lastUsage.completionTokens,
             totalTokens: lastUsage.totalTokens,
+            ...extractUsageCost(lastUsage),
           },
         }),
       }
@@ -1076,6 +1078,7 @@ export class OpenRouterTextAdapter<
               promptTokens: lastUsage.promptTokens || 0,
               completionTokens: lastUsage.completionTokens || 0,
               totalTokens: lastUsage.totalTokens || 0,
+              ...extractUsageCost(lastUsage),
             },
           }),
           finishReason,

packages/ai-openrouter/tests/cost.test.ts

@@ -0,0 +1,125 @@
+import { describe, expect, it } from 'vitest'
+import { extractUsageCost } from '../src/adapters/cost'
+
+describe('extractUsageCost', () => {
+  it('extracts a finite cost', () => {
+    expect(extractUsageCost({ cost: 0.0123 })).toEqual({ cost: 0.0123 })
+  })
+
+  it('preserves cost === 0 (not treated as absent)', () => {
+    expect(extractUsageCost({ cost: 0 })).toEqual({ cost: 0 })
+  })
+
+  it('returns empty object when cost is absent', () => {
+    expect(extractUsageCost({ promptTokens: 5 })).toEqual({})
+  })
+
+  it('returns empty object for non-number / non-finite cost', () => {
+    expect(extractUsageCost({ cost: '0.5' })).toEqual({})
+    expect(extractUsageCost({ cost: NaN })).toEqual({})
+    expect(extractUsageCost({ cost: Infinity })).toEqual({})
+    expect(extractUsageCost({ cost: null })).toEqual({})
+  })
+
+  it('returns empty object for non-object input', () => {
+    expect(extractUsageCost(undefined)).toEqual({})
+    expect(extractUsageCost(null)).toEqual({})
+    expect(extractUsageCost(42)).toEqual({})
+  })
+
+  it('reads costDetails (camelCase) alongside a valid cost', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        costDetails: { upstreamInferenceCost: 0.008 },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: 0.008 } })
+  })
+
+  it('reads cost_details (snake_case) alongside a valid cost, camelCasing keys', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        cost_details: { upstream_inference_cost: 0.008 },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: 0.008 } })
+  })
+
+  it('normalizes the full snake_case breakdown to camelCase (raw/UNKNOWN fallback path)', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.0042,
+        cost_details: {
+          upstream_inference_completions_cost: 0.0026,
+          upstream_inference_cost: 0.0038,
+          upstream_inference_prompt_cost: 0.0012,
+        },
+      }),
+    ).toEqual({
+      cost: 0.0042,
+      costDetails: {
+        upstreamInferenceCompletionsCost: 0.0026,
+        upstreamInferenceCost: 0.0038,
+        upstreamInferencePromptCost: 0.0012,
+      },
+    })
+  })
+
+  it('prefers camelCase costDetails when both are present', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        costDetails: { upstreamInferenceCost: 1 },
+        cost_details: { upstream_inference_cost: 2 },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: 1 } })
+  })
+
+  it('preserves negative detail values (e.g. cache discount)', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        costDetails: { upstreamInferenceCost: -0.002 },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: -0.002 } })
+  })
+
+  it('drops null, non-finite, and non-numeric detail entries', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        costDetails: {
+          upstreamInferenceCost: 0.5,
+          upstreamInferenceInputCost: null,
+          upstreamInferenceOutputCost: Infinity,
+          upstreamInferencePromptCost: NaN,
+          upstreamInferenceCompletionsCost: 'x',
+        },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: 0.5 } })
+  })
+
+  it('drops unknown breakdown keys', () => {
+    expect(
+      extractUsageCost({
+        cost: 0.01,
+        costDetails: {
+          upstreamInferenceCost: 0.008,
+          futureUnknownField: 0.001,
+        },
+      }),
+    ).toEqual({ cost: 0.01, costDetails: { upstreamInferenceCost: 0.008 } })
+  })
+
+  it('omits costDetails entirely when no known entries remain', () => {
+    expect(
+      extractUsageCost({ cost: 0.01, costDetails: { unknownKey: 1 } }),
+    ).toEqual({ cost: 0.01 })
+  })
+
+  it('drops an orphan costDetails when cost is absent', () => {
+    expect(
+      extractUsageCost({ costDetails: { upstreamInferenceCost: 0.008 } }),
+    ).toEqual({})
+  })
+})