Swap a Product in an Existing Clip With Omni Flash Video Edit: Cost

The short answer

Gemini Omni Flash 1.1 on Sume has a video-to-video edit mode: send video_url and a prompt through the Video Router, such as "Replace the bottle with an apple. Keep everything else the same." The docs example uses resolution: 720p. Edits bill by output second at the model's rate: $0.125 per second at 720p and $0.1875 at 1080p (catalog read 2026-10-08), so a 5-second edit at 720p is $0.625.

What the request can and cannot carry

The docs table lists video_to_video (edit) as: send video_url; the prompt gives the instructions; resolution is optional and defaults to 720p; there is no aspect_ratio or duration. video_url is the edit source, not a reference, so you cannot send it with image_url, end_image_url or reference_*_urls. Native audio is always on for this model, and the API rejects generate_audio: false.

The model range is 3 to 10 seconds for generation. The docs do not state a source-length limit for edits, so test with the length of your clip, and read capabilities from GET /v1/video-router/models for the current envelope.

Edit or regenerate

Edit when the clip is right and one object is wrong: a competitor's logo, a color, a prop. Regenerate when the camera move or the performance is wrong. A fresh text-to-video clip of 5 seconds at 720p on Omni costs the same $0.625, so the saving is not in price but in what stays. An edit keeps the framing you approved, though it may still alter details you did not name.

Check the result against the original frame by frame for drift in hands, text and edges. Use video frames to pull stills at chosen seconds from both versions for a side-by-side review; it extracts up to 24 stills per call from clips of up to 300 seconds and bills its Modal compute.

A practical test: take one 3-second clip, make the edit at 720p for $0.3750, and look at three stills. If the new object holds its shape and the shadows still fall right, run the full-length clip. If it fails on the 3-second sample, no change to the prompt is likely to save a longer one. Write the instruction as a before-and-after pair, and name the things that must not change, such as the hand position, the label text and the background.

• Do not edit footage in a way that misleads viewers about a real event or a real person.
• Keep the original clip; edits are new artifacts and the source is untouched.
• Use an Idempotency-Key per edit so a retry does not pay twice.