sume/auto defaults to 720p and 8s: 250 old prompts, pinned vs Auto

With model: "sume/auto" on POST /v1/videos, a request that sets no resolution or duration gets the Auto defaults of 720p and 8 seconds, and clips of 3 to 10 seconds at 16:9 or 9:16. Pinned to a catalog model, 250 such clips cost between $150.00 and $605.00 on Sume. Auto does not tell you which family served a request, so only a pinned model gives you a price you can compute before you submit.

OpenAI removed its Videos API on 2026-09-24 and listed no replacement. Auto is the lowest-effort replacement, and pinning is the controlled one.

What the docs say about Auto

Resolution is a pure function of the normalized request and the catalog version, so an idempotent replay gets the same price and the same route. The poll response reports sume/auto as the model, and the docs say Sume does not disclose the family and tell you not to infer it from the output. The workspace balance is reserved at submit and usage.cost shows the billable amount afterwards.

Pinned prices at the Auto defaults

For comparison, this is what 250 clips of 8 seconds, 720p and 9:16 cost on each pinned model. These are the numbers you get if you stop using Auto.

What to forecast with Auto

Because Auto's price is a function of the normalized request and the catalog version, you can forecast it from a sample. Submit 10 representative requests, read usage.cost from each completed job, and multiply the mean by 250. If the sample cost 10 x 100 cents = $10.00, the library estimate is $250.00; if the mean is 150 cents, it is $375.00.

The estimate holds only while the catalog version holds. A new catalog version can change the route and the price, so re-run the sample before a second large batch rather than reusing the first figure. A pinned model has no such drift in its own row, though its list price can still change, as the catalog notes say.

When to use which

Use Auto when the goal is a good clip and the library has no brand look to protect: prototypes, internal drafts, one-off social posts. Use a pinned model when a client expects consistent output over a campaign, when you need to forecast a bill, or when your prompts rely on a capability only some models have, such as a 12-second length or reference audio.

If you do use Auto for a library, log usage.cost for every job and compare the average with the table. Set the average next to the pinned rows: it tells you what Auto costs you in practice for your own request mix, and it needs no knowledge of the family.

• Set resolution, duration and aspect_ratio explicitly even with Auto.
• Auto rejects bitrate_mode and always generates audio on its current family; omit generate_audio.
• Pin the model you tested for any library you plan to re-render later.