Longest AI video clip in one request: 30, 15 or 10 seconds by model

Two text-to-video models on Sume reach 30 seconds in a single request: Seedance 2.5 (seedance-2.5) and Wan 3.0 (wan-3.0). The Seedance 2.0 family, Kling 3, Grok Imagine Video 1.5 and the MiniMax H3 rows stop at 15 seconds, and Gemini Omni Flash 1.1 stops at 10. Past a model's ceiling you stitch clips on a timeline instead of asking for a longer one.

Ceiling per model

Press coverage of Seedance 2.5 describes it as a 30-second model; the number that binds your request is the one in Sume's catalog, which agrees for the two 30-second rows.

What a long clip does and does not give you

A single 30-second request keeps one continuous generation, which suits a walk-through or a single-take product tour. It also costs more per attempt, so a bad take is a more expensive miss. Draft at 480p first, then rerun the winning prompt at 1080p; there is no seed on these rows, so the rerun is a new take, not a replay.

If your idea is a sequence of shots, six 5-second clips usually give you more control than one 30-second prompt, and each can be regenerated alone.

Beyond the ceiling

For more than 30 seconds, generate in segments and join them. Chain by passing the last frame of one clip as the first frame of the next where the model accepts a first frame, and keep aspect ratio and resolution identical across segments so the join does not jump.