### Check for existing issues

- [x] I searched open and closed issues before filing - [x] #18671 reports the same external symptom on v1.80.11, but it was closed as stale/not planned and the two current code paths below remain reproducible on commit `24123269ccb76f36298a2457589f08bd3141072c`

### What happened?

The proxy TPM limiter undercounts `/v1/responses` in both phases of its reservation/reconciliation flow

Before the provider call, a Responses request with non-empty `input` is classified like an embeddings request, so its output reservation is zero. The estimator also ignores the Responses API field `max_output_tokens`

After a successful provider call, the limiter receives `ResponseAPIUsage`, but `_get_total_tokens_from_usage` only handles the legacy `Usage` object or a `dict` with `prompt_tokens` and `completion_tokens`. A real `ResponseAPIUsage` object therefore reconciles as zero input, zero output, and zero total tokens

This combination can fully refund an already too-small reservation. TPM limits for keys, teams, users, and models can be under-enforced for Responses traffic, especially with concurrent requests

All values and identifiers below are synthetic and the…