sume/auto 400 unsupported_capability: 11 s, 2 s, 480p and silent audio

sume/auto returns a 400 unsupported_capability for a duration of 2 or 11 seconds, a resolution of 480p or 768p, and generate_audio: false. The request is rejected with no provider submission. Sume does not quietly reroute it to another model. The limit is the Gemini Omni Flash 1.1 envelope, which is the model sume/auto validates against for text, frame and reference inputs.

What auto accepts and refuses

Per the Videos doc:

Why a platform spec makes this matter

A spec such as TikTok's up-to-10-minutes or YouTube's up-to-3-minutes is far above what one auto clip can be. If you ask auto for 15 seconds because the ad should be 15 seconds, you hit the 400. The error is fail-closed on purpose, since the caller cannot see which model is behind sume/auto. The message and details.model name sume/auto, and supported lists the values the API accepts, so read that list instead of guessing.

Three ways to fix it

• Cut the length to 10 seconds or less and join clips in Timeline.
• Pin a model that lists the length, for example seedance-2.5 at 4 to 30 seconds.
• Leave generate_audio out of the request. If you need a silent clip, pick a model whose catalog entry has a toggle, or mute it after generation.

Reading the error

The 400 body names sume/auto rather than the model that would have served the request. That is deliberate: the caller never sees the family behind auto, so the error cannot reveal it. What you get is a list called supported with the accepted values. Parse that list and show it to whoever wrote the request, rather than retrying with a guess.

A retry loop is the wrong response. The request is wrong, not the server, so sending it again will fail the same way. Treat unsupported_capability as a validation error in your own code, log the field that was out of range, and stop.

Because auto and pinned models have different envelopes, keep them as separate code paths if you can. One path sends sume/auto with 3 to 10 seconds and 16:9 or 9:16. The other sends a named model and reads that model's lists from the catalog. Mixing them in one function is how a 15-second request ends up in the wrong place.

The tradeoff

Auto is the lowest-effort route and its envelope is narrow. Pinned models give you longer clips, but you take on reading each model's own limits. Fetch GET /v1/videos/models and check supported_durations before you submit. The 3 to 10 second limit post covers the join pattern.