Retry a 503 overload on a paid generation without a double charge

Yes. On Sume, retry a 503 with the same Idempotency-Key: the docs say to retry later with the same key unless the error says not to. Paid usage is reserved when a request is accepted and released if the work fails, so a retry is not charged on top of a failed attempt.

OpenAI's Sep 2 changelog says temporary model overload returns a 503 with the server_is_overloaded code and may include Retry-After. This page covers the Sume side, from the generation admission docs, read 2026-09-30.

What does Sume return for a 503?

A 503 with provider_capacity_exceeded or a runtime configuration error means Sume cannot start or dispatch generation work safely. The client behavior column says: retry later with the same idempotency key unless the error says not to retry.

Can a retry bill me twice?

Two mechanisms apply. First, the key: a replay of the same key returns the original job rather than a new one. Second, the money: at submit time Sume reserves the estimated amount; successful completion captures it, and failed jobs and failed queue admission release or refund the reservation where applicable. In Usage, that shows as a state.

When does reusing a key go wrong?

Only when the payload differs. The same key on a different operation or payload returns 409 idempotency_conflict, and the docs say to reuse keys only for exact retries. If you want to change the prompt, use a new key. More in idempotency keys for AI video APIs.

How should I retry?

Store the key with the request, wait (use retry-after when present), and resend the identical body with the identical key. Then check Usage for the job id: one reserved or captured row for the job means one charge. For the broader failure picture see do failed AI video jobs cost money and 429 versus 503.