21:9 banner still to video: only Seedance and MiniMax H3 list 21:9

In Sume's video catalog only the Seedance models and the two MiniMax H3 models list 21:9; Wan 3.0, Kling 3 and Gemini Omni Flash 1.1 do not. If you want a cinematic banner clip, start from a 21:9 still and send it as the first_frame to one of those.

ChatGPT Image 2.5 does not list 21:9 as a ratio preset, but its custom size rule allows it: 2352x1008 is exactly 7:3, both edges are multiples of 16, and it is under the 3:1 limit.

Making the 21:9 still

Use image_size: "2352x1008" on openai/gpt-image-2.5. That is 2,370,816 pixels, inside the 655,360 to 8,294,400 range. In the repository estimator it costs $0.0046 at low, $0.0099 at medium and $0.0414 at high from text, as estimates.

The Image API docs list 21:9 among the accepted aspect_ratio values, but support differs by model, so check GET /v1/images/models before relying on a preset.

The video side

The prices below are the SKUs the Video API catalog lists for each model. Seedance bills per thousand video tokens, so a per-second figure depends on resolution and length and is not shown here; MiniMax bills per output second.

Reading the Seedance price line

Seedance bills per thousand video tokens rather than per second, so the cost of a 5-second clip depends on resolution and the token count the model reports. Use a short 480p test to see the real cost of your prompt, and extrapolate. The cheaper tiers in the table, Seedance 2 Mini at $0.0088 and Fast at $0.0140, are the natural place to test a wide frame.

MiniMax models are easier to price on paper because they bill per output second. For a client quote that needs a number before the job runs, that predictability can matter more than the lowest rate.

• Test wide frames on the cheapest tier.
• Confirm 21:9 is in supported_aspect_ratios for your workspace.
• Keep the first frame's subject in the middle third.

Pick by budget

For a first look at motion, MiniMax H3 at 480p costs $0.3125 for 5 seconds. A full 1080p pass on H3 Max costs $1.00 for the same length. Each model lists first_frame and last_frame, so a banner can start on one still and end on another.

Check GET /v1/videos/models for your workspace before you build, since the catalog can differ by deployment. The Video API docs explain the descriptors.

If a 21:9 clip is only for a hero section on a web page, consider whether a looping 16:9 clip cropped by CSS would do. The crop is free, and Wan 3.0 or Kling 3 become available. If the clip is a deliverable on its own, such as a cinema-style ad, generate it natively at 21:9 so the model composes for the wide frame. Native composition looks better than a crop because the subject, horizon and motion are all placed for the shape you ship.

• Same ratio on still and clip avoids a crop.
• Keep text inside the middle 80 percent of the width.
• Generate at 480p first to test motion.