No seed on Sume video models: rerun a prompt reproducibly

If your Sora-era tests re-ran a prompt and expected a similar result, do not carry a seed field over to Sume. The video generation docs say no v1 model accepts seed: each model reports seed: false and rejects the field. Repeatability on Sume comes from storing the finished clip, not from re-running the prompt. Use an Idempotency-Key when you retry a submit, because a replay returns the original job rather than rendering a second take.

Three kinds of repeat, three tools

People say "reproducible" for different needs. Match each to a mechanism that exists.

What idempotency does and does not do

The docs state that a replay with the same key returns the original job. They also state that reusing a key for a different payload is an error (409 idempotency_conflict), so build keys from a stable hash of the request, not from a timestamp. A deliberate second take needs a new key, and it will bill again.

Make comparisons honest

Because nothing is seeded, a before-and-after test with one clip each proves little. Render several takes per variant, keep all of them, and judge on the set. The jobs guide shows how to read results and events for each job so you can keep a record of which model and settings made each take.

Pinning the model matters more than any seed would. With sume/auto the response echoes sume/auto and the family that ran is never disclosed, so two auto takes may come from different models.

Where this leaves your Sora prompts

OpenAI's deprecations page shows the Sora models are gone as of 2026-09-24, so there is no way to regenerate an exact Sora result. Treat old Sora files as assets to keep, and new prompts as fresh work to evaluate on the Sume model you choose.