Longest AI video clip in one API call: max seconds per Sume model

The longest single clip you can request through Sume is 30 seconds, from Seedance 2.5 and Wan 3.0. Seedance 2.0 and its Fast and Mini variants, Kling Video v3 Pro, Grok Imagine Video 1.5, MiniMax H3 and MiniMax H3 Max reach 15 seconds, and Gemini Omni Flash 1.1 reaches 10. The motion models Genjutsu and H3 Max Recast follow the source clip, up to 30 seconds. The figures come from the Sume catalog, as of 2026-10-08.

The shortest minimum is 2 seconds on Wan 3.0. Everything else starts at 3, 4 or 5.

Cost of the maximum

Maximum length is not cheap on every model. At 720p a 30-second Seedance 2.5 clip is $17.334 (648,000 tokens x $0.0214 per 1,000 x 1.25), while a 30-second Wan 3.0 clip at 720p is $3.75 (30 x $0.10 x 1.25). Wan 3.0 is the cheaper way to get a 30-second take in the catalog, and Seedance 2.5 costs about 4.6 times as much at the same resolution.

At 15 seconds the 720p-class prices are $5.67 for Seedance 2.0, $3.15 for Kling with audio, and $1.50 for H3 Max at 768p.

What the vendor pages say

The fal page for Seedance 2.5 describes native 30-second single-shot video from a single text prompt, reasoning about the whole shot at once (fal page, read 2026-10-08). The Kling Video v3 Pro page on fal describes multi-shot support within the clip (fal page, read 2026-10-08), so a 15-second Kling clip can contain more than one shot, while Seedance 2.5 is described as a single shot.

Choosing by length

• 2 to 3 seconds: Wan 3.0 (2 s) or Gemini Omni Flash 1.1 (3 s).
• 4 to 10 seconds: most models (MiniMax starts at 5); the cheapest per second are Wan 3.0 and MiniMax.
• 11 to 15 seconds: Seedance 2.0 family, Kling v3 Pro, Grok Imagine (image-to-video), MiniMax H3 or H3 Max, Wan 3.0 or Seedance 2.5.
• 16 to 30 seconds: Wan 3.0 or Seedance 2.5 only.

Going longer than the cap

For anything past 30 seconds, generate segments and join them. Sume has a timeline compose workflow for assembling clips, and each segment is an ordinary async job with its own Idempotency-Key. Plan segment boundaries so that the last frame of one clip can be the first frame of the next: Seedance, Kling, Wan 3.0, MiniMax and Gemini Omni all accept an end frame in the Sume catalog.

• Read supported_durations from GET /v1/videos/models for exact integer values.
• Send whole seconds; the catalog uses integers.
• Use callback_url for long clips rather than a polling loop.