AI video generator comparison table: price, length, sound

A useful AI video generator comparison puts every model on the same columns: what a second costs, how long one clip can run, which resolutions it makes, whether it generates sound, and what it takes as input. The per-second price alone is not enough, because the same rate can buy a different resolution, length or sound.

The table below does that for the 10 video models callable through Sume's POST /v1/videos, in alphabetical order, with no ranking. Limits come from the Video generation docs, the Sume API reference and Sume's catalog code; rates are computed from its pricing code, read 2026-09-28.

How do the AI video generators compare?

One row per model. Inputs: “first and last frame” is image-to-video, and references guide a fresh clip without being frames.

How do I compare prices when models bill differently?

Convert to the clip you will actually make. Most models bill per second of output at a rate set by resolution; kling-3 sets its rate by sound instead. The Seedance ids bill per 1,000 video tokens, counted from the frame's pixels and the clip's length, so the table also shows a 720p 16:9 second for each.

Then multiply by the length you need, and add the 5.5% agent fee. How much does AI video cost per second? works through 10-second and one-minute examples.

• Send resolution explicitly. With none, wan-3.0 reserves at its 1080p rate in current code.
• The two MiniMax models and Gemini Omni Flash 1.1 always make sound, so their rates include it.
• A failed job's reserve is refunded; see Do failed AI video jobs cost money?

Which AI video generator should I pick?

Pick by the constraint your clip has to meet:

• One clip over 15 seconds: seedance-2.5 or wan-3.0, both up to 30 seconds.
• 4K: gemini-omni-flash-1.1 lists 4K among its resolutions; minimax-h3 prices 2K and 4K as upscales of its native 768p.
• Editing a clip you already have: gemini-omni-flash-1.1, the only model here with a video edit mode.
• Audio references: the Seedance ids, wan-3.0 and the two MiniMax models.
• One still animated with no sound: grok-imagine-video-1.5.
• Veo or Sora: not in Sume's catalog. OpenAI's deprecations page lists the Videos API, sora-2 and sora-2-pro with a shutdown date of 2026-09-24; Sora API alternatives covers the move.

What does this table leave out?

• sume/auto: Sume picks the model and does not disclose which one ran, so it has no row.
• Quality, look and render time: no row states them, because none of the sources publishes a comparison. Test your own prompt on the models that pass your constraints.
• Per-model details such as reference counts and aspect ratios: GET /v1/videos/models lists each model's supported_* fields, and AI video length limits by model covers the length errors.