diff --git a/apps/web/src/lib/ai-gateway/finishReason.test.ts b/apps/web/src/lib/ai-gateway/finishReason.test.ts
new file mode 100644
index 000000000..8ec6f0cd7
--- /dev/null
+++ b/apps/web/src/lib/ai-gateway/finishReason.test.ts
@@ -0,0 +1,38 @@
+import {
+  ERROR_FINISH_REASONS,
+  NON_ERROR_FINISH_REASONS,
+  isErrorFinishReason,
+} from '@/lib/ai-gateway/finishReason';
+
+describe('finishReason', () => {
+  it('classifies known error finish_reasons as errors', () => {
+    for (const reason of ERROR_FINISH_REASONS) {
+      expect(isErrorFinishReason(reason)).toBe(true);
+    }
+  });
+
+  it('classifies known non-error finish_reasons as non-errors', () => {
+    for (const reason of NON_ERROR_FINISH_REASONS) {
+      expect(isErrorFinishReason(reason)).toBe(false);
+    }
+  });
+
+  it('treats null/undefined as non-error', () => {
+    expect(isErrorFinishReason(null)).toBe(false);
+    expect(isErrorFinishReason(undefined)).toBe(false);
+  });
+
+  it('treats unrecognised string values as non-error', () => {
+    // Unknown values should not flip hasError; other signals (statusCode,
+    // wasAborted) handle those cases. This also keeps us from creating
+    // spurious error rows when a new provider adds a new stop reason.
+    expect(isErrorFinishReason('something_new_from_provider')).toBe(false);
+  });
+
+  it('does not double-count any reason in both lists', () => {
+    const intersection = NON_ERROR_FINISH_REASONS.filter(r =>
+      (ERROR_FINISH_REASONS as readonly string[]).includes(r)
+    );
+    expect(intersection).toEqual([]);
+  });
+});
diff --git a/apps/web/src/lib/ai-gateway/finishReason.ts b/apps/web/src/lib/ai-gateway/finishReason.ts
new file mode 100644
index 000000000..27706f602
--- /dev/null
+++ b/apps/web/src/lib/ai-gateway/finishReason.ts
@@ -0,0 +1,40 @@
+// The two lists below enumerate every distinct value observed for the
+// `finish_reason` column in production `microdollar_usage` logs.
+// `unknown` / `other` are kept as non-error catch-alls so a novel upstream
+// value does not immediately inflate the error rate.
+
+export const NON_ERROR_FINISH_REASONS = [
+  'stop',
+  'tool_calls',
+  'tool-calls',
+  'end_turn',
+  'completed',
+  'tool_use',
+  'stop_sequence',
+  'unknown',
+  'other',
+] as const;
+
+export const ERROR_FINISH_REASONS = [
+  'length',
+  'max_tokens',
+  'content_filter',
+  'content-filter',
+  'error',
+  'network_error',
+  'failed',
+  'model_context_window_exceeded',
+  'engine_overloaded',
+  'refusal',
+  'incomplete',
+  'in_progress',
+] as const;
+
+const errorFinishReasonSet: ReadonlySet<string> = new Set(ERROR_FINISH_REASONS);
+
+// `null` / `undefined` return false: an absent finish_reason is handled by
+// the `wasAborted` / `reportedError` signals in the parsers, not here.
+export function isErrorFinishReason(finish_reason: string | null | undefined): boolean {
+  if (finish_reason == null) return false;
+  return errorFinishReasonSet.has(finish_reason);
+}
diff --git a/apps/web/src/lib/ai-gateway/processUsage.messages.ts b/apps/web/src/lib/ai-gateway/processUsage.messages.ts
index 67af35004..dc343e464 100644
--- a/apps/web/src/lib/ai-gateway/processUsage.messages.ts
+++ b/apps/web/src/lib/ai-gateway/processUsage.messages.ts
@@ -17,6 +17,7 @@ import {
   drainSseStream,
   extractVercelIsByok,
 } from '@/lib/ai-gateway/processUsage.shared';
+import { isErrorFinishReason } from '@/lib/ai-gateway/finishReason';
 import type Anthropic from '@anthropic-ai/sdk';
 
 type MaybeHasVercelProviderMetadata = {
@@ -170,7 +171,7 @@ export async function parseMessagesMicrodollarUsageFromStream(
 
   const coreProps = {
     messageId,
-    hasError: reportedError || wasAborted,
+    hasError: reportedError || wasAborted || isErrorFinishReason(finish_reason),
     model,
     responseContent,
     inference_provider,
@@ -205,14 +206,15 @@ export function parseMessagesMicrodollarUsageFromString(
     .map(c => c.text)
     .join('');
 
+  const finish_reason = responseJson?.stop_reason ?? null;
   const coreProps = {
     messageId: responseJson?.id ?? null,
-    hasError: !responseJson?.model || statusCode >= 400,
+    hasError: !responseJson?.model || statusCode >= 400 || isErrorFinishReason(finish_reason),
     model: responseJson?.model ?? null,
     responseContent,
     inference_provider,
     upstream_id: null,
-    finish_reason: responseJson?.stop_reason ?? null,
+    finish_reason,
     latency: null,
     moderation_latency: null,
     generation_time: null,
diff --git a/apps/web/src/lib/ai-gateway/processUsage.responses.ts b/apps/web/src/lib/ai-gateway/processUsage.responses.ts
index eb8e2eae3..b5fefef3e 100644
--- a/apps/web/src/lib/ai-gateway/processUsage.responses.ts
+++ b/apps/web/src/lib/ai-gateway/processUsage.responses.ts
@@ -18,6 +18,7 @@ import {
   drainSseStream,
   extractVercelIsByok,
 } from '@/lib/ai-gateway/processUsage.shared';
+import { isErrorFinishReason } from '@/lib/ai-gateway/finishReason';
 
 // OpenRouter adds cost fields to the standard Responses API usage object.
 // ref: https://openrouter.ai/docs/use-cases/usage-accounting#response-format
@@ -204,7 +205,7 @@ export async function parseResponsesMicrodollarUsageFromStream(
 
   const coreProps = {
     messageId,
-    hasError: reportedError || wasAborted,
+    hasError: reportedError || wasAborted || isErrorFinishReason(finish_reason),
     model,
     responseContent,
     inference_provider,
diff --git a/apps/web/src/lib/ai-gateway/processUsage.ts b/apps/web/src/lib/ai-gateway/processUsage.ts
index 1a79696f9..d001dd753 100644
--- a/apps/web/src/lib/ai-gateway/processUsage.ts
+++ b/apps/web/src/lib/ai-gateway/processUsage.ts
@@ -54,6 +54,7 @@ import {
   parseMessagesMicrodollarUsageFromString,
 } from '@/lib/ai-gateway/processUsage.messages';
 import { OPENROUTER_BYOK_COST_MULTIPLIER } from '@/lib/ai-gateway/processUsage.constants';
+import { isErrorFinishReason } from '@/lib/ai-gateway/finishReason';
 import {
   computeOpenRouterCostFields,
   drainSseStream,
@@ -785,7 +786,7 @@ export async function parseMicrodollarUsageFromStream(
   const coreProps = {
     kiloUserId,
     messageId,
-    hasError: reportedError || wasAborted,
+    hasError: reportedError || wasAborted || isErrorFinishReason(finish_reason),
     model,
     responseContent,
     inference_provider,
@@ -822,10 +823,11 @@ export function parseMicrodollarUsageFromString(
     });
   }
   const choice = responseJson?.choices?.[0];
+  const finish_reason = choice?.finish_reason ?? null;
   const coreProps = {
     kiloUserId,
     messageId: responseJson?.id ?? null,
-    hasError: !responseJson?.model || statusCode >= 400,
+    hasError: !responseJson?.model || statusCode >= 400 || isErrorFinishReason(finish_reason),
     model: responseJson?.model ?? null,
     responseContent: choice?.message.content ?? '',
     inference_provider:
@@ -833,7 +835,7 @@ export function parseMicrodollarUsageFromString(
       choice?.message?.provider_metadata?.gateway?.routing?.finalProvider ??
       null,
     upstream_id: null,
-    finish_reason: choice?.finish_reason ?? null,
+    finish_reason,
     latency: null,
     moderation_latency: null,
     generation_time: null,