Token usage mismatch when stopping generation mid-stream

[398168441]

What happened?

When I click Stop during streaming, the values passed to spendTokens.tokenUsage.promptTokens and .completionTokens don’t match the tokens actually charged by OpenRouter.

Version Information

When I click Stop during streaming, the values passed to spendTokens.tokenUsage.promptTokens and .completionTokens don’t match the tokens actually charged by OpenRouter.

Steps to Reproduce
1. Send a prompt.
2. While the model is streaming, click Stop.
3. Compare tokenUsage from spendTokens with OpenRouter’s reported usage.

Expected: tokenUsage.promptTokens and completionTokens match OpenRouter’s actual usage.
Actual: They do not match.

Steps to Reproduce

Steps to Reproduce
1. Send a prompt.
2. While the model is streaming, click Stop.
3. Compare tokenUsage from spendTokens with OpenRouter’s reported usage.

What browsers are you seeing the problem on?

Chrome

Relevant log output

Expected: tokenUsage.promptTokens and completionTokens match OpenRouter’s actual usage.
Actual: They do not match.

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Token metadata is not available from the provider when cancelling a request since the connection is severed, so an estimate is made instead.

[398168441]

I found that the estimated value and the actual token usage sometimes differ significantly.

[398168441]

I’ve tested this many times — when I manually stop the request, the promptTokens calculated here aren’t just slightly off; they’re flat-out wrong.