Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

account for large partial stream usage

Louis Beaumontfc89fe3f4923
diff --git a/packages/ai-gateway/src/services/cost-tracker.ts b/packages/ai-gateway/src/services/cost-tracker.ts
index fd0949ba9..aa5d0c2f1 100644
--- a/packages/ai-gateway/src/services/cost-tracker.ts
+++ b/packages/ai-gateway/src/services/cost-tracker.ts
@@ -197,7 +197,15 @@ export function getStreamModelCost(
cache?: CacheUsage,
): number {
if (inputTokens <= 0 || outputTokens <= 0) {
- return getModelCost(model, null, null);
+ // Preserve any large partial counter already observed. Falling all the way
+ // back to the average would undercount a cancelled 100k-token prompt just
+ // because its final output usage event never arrived.
+ return getModelCost(
+ model,
+ Math.max(inputTokens, DEFAULT_INPUT_TOKENS),
+ Math.max(outputTokens, DEFAULT_OUTPUT_TOKENS),
+ cache,
+ );
}
return getModelCost(model, inputTokens, outputTokens, cache);
}
diff --git a/packages/ai-gateway/src/test/cost-tracker.unit.test.ts b/packages/ai-gateway/src/test/cost-tracker.unit.test.ts
index 5a3e4ba98..43108f534 100644
--- a/packages/ai-gateway/src/test/cost-tracker.unit.test.ts
+++ b/packages/ai-gateway/src/test/cost-tracker.unit.test.ts
@@ -20,6 +20,12 @@ describe('getModelCost — cache-aware pricing', () => {
getModelCost('gpt-5.6-sol', null, null),
);
expect(getStreamModelCost('gpt-5.6-sol', 0, 0)).toBeGreaterThan(0);
+ expect(getStreamModelCost('gpt-5.6-sol', 100_000, 0)).toBe(
+ getModelCost('gpt-5.6-sol', 100_000, 500),
+ );
+ expect(getStreamModelCost('gpt-5.6-sol', 100_000, 0)).toBeGreaterThan(
+ getModelCost('gpt-5.6-sol', null, null),
+ );
});
it('uses the exact Opus 5 standard price and cache rates', () => {
const uncached = getModelCost('claude-opus-5', 100_000, 1_000);