Gemini Omni Flash 1.1 from a still: 16:9 or 9:16 only, cost 360p to 4K

Gemini Omni Flash 1.1 on Sume lists only 16:9 and 9:16, so a square or 4:3 still is not a valid handoff for it; make the still in one of those two ratios. A 5-second clip costs $0.1875 at 360p, $0.6250 at 720p, $0.9375 at 1080p and $1.88 at 4K.

The model takes first_frame and last_frame, accepts image_url and video_url input references, and generates audio. Its duration range is 3 to 10 seconds.

Price by resolution

Per-second SKUs from the Video API catalog, and the cost of common clip lengths.

Match the still to the clip

Make a 16:9 still with aspect_ratio: "16:9" or a 9:16 one with aspect_ratio: "9:16", both of which the Image API lists for ChatGPT Image 2.5. Then pass the URL in frame_images with frame_type: "first_frame" and set the same aspect_ratio on the video request.

If your source is 1:1 or 4:5, either generate a new still at the right ratio or use a model that lists your ratio; the table in the 4:5 post shows which models take what.

Duration limits and the cost of a retry

The 3 to 10 second range keeps a single clip cheap but means a longer sequence is several jobs. A 30-second spot is three 10-second clips, and at 1080p each costs $1.88, so the sequence costs $5.63. Cutting it from clips shot at 720p brings it to $3.75.

Since a retry bills like a new job, the 360p draft is the cheapest insurance you can buy. Run every new idea there once. A rejected idea costs a few cents and an approved one costs the full price only once.

• Use the same still for several clips by changing only the prompt.
• Use last_frame to make clips join smoothly.
• Keep an eye on the 4K tier; it is twice the 1080p rate.

Use 360p for the draft

The 360p tier costs one fifth of the 1080p rate per second, which makes it the right place to test a prompt: ten 5-second drafts cost $1.88 against $9.38 at 1080p. Move up only once the motion is right.

Read supported_resolutions and supported_aspect_ratios from GET /v1/videos/models before launching a batch. See the Video API docs.

Write the ratio, resolution and duration into the request explicitly and read them back from the job result. A default that differs from what you assumed is the most common source of a clip that arrives in the wrong shape.