AI video API clip length: which Sume models accept it, and the cost

Each video model on Sume accepts a window of clip lengths, published as supported_durations at GET /v1/videos/models, and refuses a request outside it at validation. The length you need decides which models you can use; price decides between them.

Worked example: 12-second AI video API: nine Sume models that accept 12 seconds

A 12-second request is accepted by 9 of the ten text-or-image video rows in Sume's catalog, and refused by the rest with a 400 at validation. Sume advertises each row's supported_durations at GET /v1/videos/models, and the Video Router docs list the same windows. Twelve seconds is just past Gemini Omni Flash 1.1's 10-second ceiling, which makes it a clean dividing line.

Which Sume models accept 12 seconds?

Accepted: Seedance 2.5 (seedance-2.5), Seedance 2.0 (seedance-2), Seedance 2.0 Fast (seedance-2-fast), Seedance 2.0 Mini (seedance-2-mini), Kling Video v3 Pro (kling-3), Wan 3.0 (wan-3.0), Grok Imagine Video 1.5 (grok-imagine-video-1.5), MiniMax H3 (minimax-h3), MiniMax H3 Max (minimax-h3-max).

Refused: gemini-omni-flash-1.1. For H3 Max Recast and Higgsfield Genjutsu the length is not a choice at all, since the output keeps the length of the source video.

• Per-row windows, in whole seconds: seedance-2.5 (4 to 30 s), seedance-2 (4 to 15 s), seedance-2-fast (4 to 15 s), seedance-2-mini (4 to 15 s), kling-3 (4 to 15 s), wan-3.0 (2 to 30 s), grok-imagine-video-1.5 (4 to 15 s), minimax-h3 (5 to 15 s), minimax-h3-max (5 to 15 s), gemini-omni-flash-1.1 (3 to 10 s).

What does a 12-second clip cost?

Billable cost is the per-second rate times the seconds, rounded up to the cent, where the rate is the provider list times 1.25. The table prices only the rows with a published per-second rate; the Seedance rows bill per video token, so read their cost from the finished job rather than from this table.

The lowest and highest resolution columns bracket the range: resolution is the biggest lever on any per-second row, so a draft at the lowest tier and a final at the highest is the cheapest way to iterate.

How is a 12-second job billed and settled?

Sume reserves the workspace balance on submit at the provider list times 1.25 and reports the final billable amount as usage.cost on the poll response at GET /v1/videos/{id}. The reserve is an estimate from the request you sent, so a longer duration reserves more up front. If you send an Idempotency-Key, a retry of the same body returns the original job instead of creating and billing a second one.

Poll until the job reaches a terminal status, then download with GET /v1/videos/{id}/content?index=0. A content request on a failed job is a 409 job_failed, so check the status first. The Jobs and results page covers the lifecycle.

How do I request 12 seconds?

Send duration: 12 as a whole number on POST /v1/videos. Do not send a fractional value; the windows are in whole seconds. If you would rather not pick a row by hand, GET /v1/videos/models returns supported_durations, and filtering it for 12 is a few lines, as shown in the Python filter post.

Remember that sume/auto only validates 3 to 10 seconds, so a request outside that window needs a pinned model.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"seedance-2.5","prompt":"A slow dolly past a rain-streaked tram window","duration":12}'

Every clip length in one table

This page replaces 7 posts that asked the same question with different numbers. Each row is one of them, with the answer it gave when it was published.