feat(ai-gateway): log Vercel upstream provider id as upstream_id

apps/web/src/lib/ai-gateway/processUsage.messages.ts

@@ -16,6 +16,7 @@ import {
   computeVercelCostMicrodollars,
   drainSseStream,
   extractVercelIsByok,
+  extractVercelUpstreamId,
 } from '@/lib/ai-gateway/processUsage.shared';
 import { isErrorFinishReason } from '@/lib/ai-gateway/finishReason';
 import type Anthropic from '@anthropic-ai/sdk';
@@ -222,7 +223,7 @@ export async function parseMessagesMicrodollarUsageFromStream(
     responseContent,
     inference_provider,
     finish_reason,
-    upstream_id: null,
+    upstream_id: extractVercelUpstreamId(providerMetadata),
     latency: null,
     moderation_latency: null,
     generation_time: null,
@@ -260,7 +261,7 @@ export function parseMessagesMicrodollarUsageFromString(
     model: getVercelResolvedModel(providerMetadata) ?? responseJson?.model ?? null,
     responseContent,
     inference_provider,
-    upstream_id: null,
+    upstream_id: extractVercelUpstreamId(providerMetadata),
     finish_reason,
     latency: null,
     moderation_latency: null,

apps/web/src/lib/ai-gateway/processUsage.responses.ts

@@ -17,6 +17,7 @@ import {
   computeVercelCostMicrodollars,
   drainSseStream,
   extractVercelIsByok,
+  extractVercelUpstreamId,
 } from '@/lib/ai-gateway/processUsage.shared';
 import { isErrorFinishReason } from '@/lib/ai-gateway/finishReason';
 
@@ -210,7 +211,7 @@ export async function parseResponsesMicrodollarUsageFromStream(
     responseContent,
     inference_provider,
     finish_reason,
-    upstream_id: null,
+    upstream_id: extractVercelUpstreamId(providerMetadata),
     latency: null,
     moderation_latency: null,
     generation_time: null,
@@ -240,7 +241,7 @@ export function parseResponsesMicrodollarUsageFromString(
     model: responseJson?.model ?? null,
     responseContent: responseJson?.output ? extractResponseContent(responseJson.output) : '',
     inference_provider,
-    upstream_id: null,
+    upstream_id: extractVercelUpstreamId(providerMetadata),
     finish_reason: responseJson?.status ?? null,
     latency: null,
     moderation_latency: null,

apps/web/src/lib/ai-gateway/processUsage.shared.ts

@@ -5,6 +5,7 @@ import { errorExceptInTest } from '@/lib/utils.server';
 import { OPENROUTER_BYOK_COST_MULTIPLIER } from '@/lib/ai-gateway/processUsage.constants';
 import type {
   NotYetCostedUsageStats,
+  VercelProviderAttempt,
   VercelProviderMetaData,
 } from '@/lib/ai-gateway/processUsage.types';
 
@@ -62,6 +63,22 @@ export function computeVercelCostMicrodollars(
   return toMicrodollars(isNaN(marketCost_USD) ? 0 : marketCost_USD);
 }
 
+/**
+ * Finds the provider attempt that served the request in the Vercel AI Gateway
+ * routing metadata: the successful provider attempt within the successful model
+ * attempt, falling back to the last attempt of each when none reports success.
+ */
+function findVercelServingProviderAttempt(
+  vercelGateway: NonNullable<VercelProviderMetaData['gateway']> | undefined | null
+): VercelProviderAttempt | undefined {
+  const modelAttempts = vercelGateway?.routing?.modelAttempts;
+  if (!modelAttempts) return undefined;
+  const servingModel = modelAttempts.find(m => m.success) ?? modelAttempts.at(-1);
+  const providerAttempts = servingModel?.providerAttempts;
+  if (!providerAttempts) return undefined;
+  return providerAttempts.find(p => p.success) ?? providerAttempts.at(-1);
+}
+
 /**
  * Extracts whether the Vercel AI Gateway served the request with BYOK credentials.
  *
@@ -73,18 +90,27 @@ export function computeVercelCostMicrodollars(
 export function extractVercelIsByok(
   vercelGateway: NonNullable<VercelProviderMetaData['gateway']> | undefined | null
 ): boolean | null {
-  const modelAttempts = vercelGateway?.routing?.modelAttempts;
-  if (!modelAttempts) return null;
-  const successfulModel = modelAttempts.find(m => m.success) ?? modelAttempts.at(-1);
-  const providerAttempts = successfulModel?.providerAttempts;
-  if (!providerAttempts) return null;
-  const successfulProvider = providerAttempts.find(p => p.success) ?? providerAttempts.at(-1);
-  const credentialType = successfulProvider?.credentialType;
+  const credentialType = findVercelServingProviderAttempt(vercelGateway)?.credentialType;
   if (credentialType === 'byok') return true;
   if (credentialType === 'system') return false;
   return null;
 }
 
+/**
+ * Extracts the upstream provider's own request/response id from the Vercel AI
+ * Gateway routing metadata, so it can be stored as
+ * `microdollar_usage_metadata.upstream_id` without a `/generation` lookup.
+ *
+ * Mirrors the attempt selection of `extractVercelIsByok`. Returns `null` when
+ * the gateway did not report an id.
+ */
+export function extractVercelUpstreamId(
+  vercelProviderMetadata: VercelProviderMetaData | undefined | null
+): string | null {
+  const attempt = findVercelServingProviderAttempt(vercelProviderMetadata?.gateway);
+  return attempt?.providerRequestId ?? attempt?.providerResponseId ?? null;
+}
+
 export function isResponseInterruptedError(error: unknown): boolean {
   if (typeof error !== 'object' || error === null || !('name' in error)) return false;
   return error.name === 'ResponseAborted' || error.name === 'TimeoutError';

apps/web/src/lib/ai-gateway/processUsage.test.ts

@@ -296,6 +296,96 @@ describe('parseMicrodollarUsageFromStream approval tests', () => {
     expect(result.hasError).toBe(true);
     expect(result.status_code).toBe(502);
   });
+
+  test('records the Vercel upstream provider request id as upstream_id', async () => {
+    const chunk = `data: {"id":"gen_01KYMFY57BAFKZGK45197SCRGD","object":"chat.completion.chunk","created":1785246720,"model":"moonshotai/kimi-k3-fast","choices":[{"index":0,"delta":{"provider_metadata":{"fireworks":{},"gateway":{"routing":{"originalModelId":"moonshotai/kimi-k3-fast","resolvedProvider":"fireworks","fallbacksAvailable":[],"canonicalSlug":"moonshotai/kimi-k3-fast","finalProvider":"fireworks","speed":"fast","modelAttemptCount":1,"modelAttempts":[{"canonicalSlug":"moonshotai/kimi-k3-fast","success":true,"providerAttemptCount":1,"providerAttempts":[{"provider":"fireworks","credentialType":"system","success":true,"startTime":1785246717210,"endTime":1785246721497,"providerRequestId":"chatcmpl-ccebc94c526d4f8797cfe00023478a9a","statusCode":200,"providerResponseId":"chatcmpl-ccebc94c526d4f8797cfe00023478a9a"}]}],"totalProviderAttemptCount":1},"cost":"0.0474822","marketCost":"0.0474822","generationId":"gen_01KYMFY57BAFKZGK45197SCRGD"}}},"logprobs":null,"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":91904,"completion_tokens":134,"total_tokens":92038,"cost":0.0474822,"is_byok":false,"prompt_tokens_details":{"cached_tokens":91136},"cost_details":{"upstream_inference_cost":null},"completion_tokens_details":{"reasoning_tokens":0}},"generationId":"gen_01KYMFY57BAFKZGK45197SCRGD"}\n\ndata: [DONE]\n\n`;
+
+    const stream = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode(chunk));
+        controller.close();
+      },
+    });
+
+    const result = await parseMicrodollarUsageFromStream(
+      stream,
+      'fake-user-id',
+      undefined,
+      'vercel',
+      200
+    );
+
+    expect(result.upstream_id).toBe('chatcmpl-ccebc94c526d4f8797cfe00023478a9a');
+    expect(result.inference_provider).toBe('fireworks');
+    expect(result.cost_mUsd).toBe(47482);
+  });
+
+  test('leaves upstream_id null when the Vercel gateway reports no provider request id', async () => {
+    const chunk = `data: {"id":"gen-1","object":"chat.completion.chunk","created":1,"model":"moonshotai/kimi-k3-fast","choices":[{"index":0,"delta":{"provider_metadata":{"gateway":{"routing":{"finalProvider":"fireworks","modelAttempts":[{"success":true,"providerAttempts":[{"provider":"fireworks","credentialType":"system","success":true}]}]},"marketCost":"0.001"}}},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}\n\ndata: [DONE]\n\n`;
+
+    const stream = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode(chunk));
+        controller.close();
+      },
+    });
+
+    const result = await parseMicrodollarUsageFromStream(
+      stream,
+      'fake-user-id',
+      undefined,
+      'vercel',
+      200
+    );
+
+    expect(result.upstream_id).toBeNull();
+  });
+
+  test('records the Vercel upstream provider request id for non-streamed responses', () => {
+    const response = JSON.stringify({
+      id: 'gen-2',
+      object: 'chat.completion',
+      created: 1,
+      model: 'moonshotai/kimi-k3-fast',
+      choices: [
+        {
+          index: 0,
+          finish_reason: 'stop',
+          message: {
+            role: 'assistant',
+            content: 'hi',
+            provider_metadata: {
+              gateway: {
+                routing: {
+                  finalProvider: 'fireworks',
+                  modelAttempts: [
+                    {
+                      success: true,
+                      providerAttempts: [
+                        {
+                          provider: 'fireworks',
+                          credentialType: 'system',
+                          success: true,
+                          providerResponseId: 'chatcmpl-non-streamed',
+                        },
+                      ],
+                    },
+                  ],
+                },
+                marketCost: '0.001',
+              },
+            },
+          },
+        },
+      ],
+      usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+    });
+
+    const result = parseMicrodollarUsageFromString(response, 'fake-user-id', 200);
+
+    expect(result.upstream_id).toBe('chatcmpl-non-streamed');
+    expect(result.inference_provider).toBe('fireworks');
+  });
 });
 
 const sampleReqDir = join(process.cwd(), 'src/tests/req_sample');

apps/web/src/lib/ai-gateway/processUsage.ts

@@ -76,6 +76,7 @@ import {
   computeOpenRouterCostFields,
   drainSseStream,
   extractVercelIsByok,
+  extractVercelUpstreamId,
   isResponseInterruptedError,
 } from '@/lib/ai-gateway/processUsage.shared';
 import {
@@ -1119,7 +1120,7 @@ export async function parseMicrodollarUsageFromStream(
     responseContent,
     inference_provider,
     finish_reason,
-    upstream_id: null,
+    upstream_id: extractVercelUpstreamId(vercelProviderMetadata),
     latency: null,
     moderation_latency: null,
     generation_time: null,
@@ -1152,17 +1153,16 @@ export function parseMicrodollarUsageFromString(
   }
   const choice = responseJson?.choices?.[0];
   const finish_reason = choice?.finish_reason ?? null;
+  const vercelProviderMetadata = choice?.message?.provider_metadata ?? null;
   const coreProps = {
     kiloUserId,
     messageId: responseJson?.id ?? null,
     hasError: !responseJson?.model || statusCode >= 400 || isErrorFinishReason(finish_reason),
     model: responseJson?.model ?? null,
     responseContent: choice?.message.content ?? '',
     inference_provider:
-      responseJson?.provider ??
-      choice?.message?.provider_metadata?.gateway?.routing?.finalProvider ??
-      null,
-    upstream_id: null,
+      responseJson?.provider ?? vercelProviderMetadata?.gateway?.routing?.finalProvider ?? null,
+    upstream_id: extractVercelUpstreamId(vercelProviderMetadata),
     finish_reason,
     latency: null,
     moderation_latency: null,
@@ -1172,11 +1172,7 @@ export function parseMicrodollarUsageFromString(
     status_code: statusCode,
   };
 
-  const costs = processOpenRouterUsage(
-    responseJson?.usage,
-    coreProps,
-    choice?.message?.provider_metadata ?? null
-  );
+  const costs = processOpenRouterUsage(responseJson?.usage, coreProps, vercelProviderMetadata);
 
   return { ...coreProps, ...costs };
 }
@@ -1252,6 +1248,7 @@ export async function processTokenData(
     }
 
     genStats.model = usageStats.model; // openrouter bug?
+    genStats.upstream_id ??= usageStats.upstream_id; // keep the id the response already reported
     genStats.hasError = usageStats.hasError; // retain by choice
     genStats.status_code = usageStats.status_code; // retain by choice
     genStats.streamed ??= usageContext.isStreaming;

apps/web/src/lib/ai-gateway/processUsage.types.ts

@@ -23,6 +23,10 @@ export type VercelProviderAttempt = {
   provider?: string;
   credentialType?: string;
   success?: boolean;
+  /** The upstream provider's own request id, e.g. Fireworks' `chatcmpl-…`. */
+  providerRequestId?: string;
+  /** The upstream provider's own response id; usually equal to `providerRequestId`. */
+  providerResponseId?: string;
 };
 
 export type VercelModelAttempt = {

apps/web/src/tests/sample/vercel-messages.log.resp.sse.approved.json

@@ -5,7 +5,7 @@
   "responseContent": "Hi there! 👋 How are you doing? Is there something I can help you with today?",
   "inference_provider": "bedrock",
   "finish_reason": "end_turn",
-  "upstream_id": null,
+  "upstream_id": "305d0d3d-1a2c-47fe-a99f-96f30082adbd",
   "latency": null,
   "moderation_latency": null,
   "generation_time": null,

apps/web/src/tests/sample/vercel-responses.log.resp.json.approved.json

@@ -4,7 +4,7 @@
   "model": "gpt-5.4-2026-03-05",
   "responseContent": "hi",
   "inference_provider": "openai",
-  "upstream_id": null,
+  "upstream_id": "req_869e7f2d33124834a2d0a7fb0a271c1a",
   "finish_reason": "completed",
   "latency": null,
   "moderation_latency": null,

apps/web/src/tests/sample/vercel-responses.log.resp.sse.approved.json

@@ -5,7 +5,7 @@
   "responseContent": "hi",
   "inference_provider": "openai",
   "finish_reason": "completed",
-  "upstream_id": null,
+  "upstream_id": "req_869e7f2d33124834a2d0a7fb0a271c1a",
   "latency": null,
   "moderation_latency": null,
   "generation_time": null,