OpenAI image-encoding fix: rerun workflows with the right key

If you decide to rerun work after OpenAI's image-encoding fix, a Sume request sent again with the same Idempotency-Key returns the original instead of running again. To actually redo a run, send a new key. Sume's docs say retrying a failed Format run takes a new Idempotency-Key, because the old one is bound to the receipt you already have.

The OpenAI note is from its changelog (Sep 25 entry, read 2026-10-01). It concerns OpenAI's GPT-6 models, not a Sume change; the key rules are from Formats errors.

What did OpenAI say to do?

The Sep 25 entry says OpenAI fixed a bug in image encoding that degraded image understanding in GPT-6 Sol and GPT-6 Luna. It recommends that, if your use cases involve image inputs, you rerun your evaluations and retry workflows affected by the issue. Which of your workflows were affected is for you to decide from your own logs.

Which key do I send when I rerun?

It depends on whether you want the old result or a new one.

Does a replay cost anything?

For Format runs, the docs say a 4xx at create, an idempotent 200 replay, and a skipped run cost nothing. They also say generation that finished before a cancel or a failure is billed, and that a later step failing does not refund it. So a rerun with a new key is new work and new spend; compare usage.billable_amount_usd_micros with your spend cap first.

How do I avoid redoing finished clips?

When a failed run left clips behind, the docs prefer continuing the run over a fresh one, because finished clips on the thread are not regenerated. See failed run retry needs a new key, or continue.