Draft with sume/auto, finish on a pinned model: why the look differs

Do not draft with sume/auto and then finish on a pinned model like gemini-omni-flash-1.1 if the draft is meant to preview the look. Sume does not disclose which family served an Auto request, so you cannot reproduce that model for the final. A draft is only a preview when it uses the same catalog id as the final, at a lower resolution.

What the docs say about Auto

These points come from Sume docs: Video generation and Video Router, both read 2026-10-05.

• sume/auto lets Sume select the model family. The poll response reports "model": "sume/auto".
• Sume does not disclose which family served the request, and the docs say not to infer it from the output.
• 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 Auto create controls default to 720p and 8 seconds, with 3 to 10 second clips at 16:9 or 9:16.

Where this bites

Google's 360p draft mode exists to prototype and iterate on a storyboard, at a third of the cost of its 720p mode (Google, read 2026-10-05). The idea only works when the draft and the final come from one model, because the draft answers the question: will this prompt produce this motion? If an Auto draft ran on another family, a good draft says nothing about the pinned final.

The reverse also holds. An Auto final after a pinned draft can land on a model that never saw your draft. Auto may suit one-off requests where you accept the model's look. It is not a draft tier.

A rule you can code

Pick the model once per shot, store it with the shot, and let only the resolution change.

Check the catalog first

Before you hard-code the id, read GET /v1/videos/models. It lists supported_resolutions, supported_durations and pricing_skus for each id, so a typo or a retired id fails early. Sume tells you the catalog changes over time, which is why the id belongs in config and not in the prompt text. If you do want Sume to pick the model on a future request, send sume/auto as a deliberate choice for that shot.

When Auto is the right choice

Auto is built for the case where you care about the clip and not the model. A catalog that grows through the year, such as the 2026 releases of Seedance 2.5 and Wan 3.0, can change what Auto picks without a code change on your side. That is a feature for a throwaway social clip and a risk for a series that must keep a look.

A short test settles your own case. Generate the same prompt twice on sume/auto and twice on a pinned id, and compare the pairs. If you cannot tell the Auto pair from the pinned pair, Auto is fine for that prompt family. If the pinned pair is more alike, pin.