Omni edit loop on Sume: three instructions, three jobs, about $3.00
Editing a clip with Gemini Omni one instruction at a time is one Sume job per pass. If each pass returns 8 s at 720p, three passes cost $3.00.

On Sume each Omni edit instruction is its own job, so a three-instruction edit loop is three jobs. If every pass returns 8 seconds at 720p, each costs 8 x $0.125 = $1.00 and the loop costs $3.00. Google describes Gemini Omni Flash as a model for video generation and conversational video editing; on Sume the edit mode is the video_to_video capability of gemini-omni-flash-1.1, reached by sending a video_url.
How an edit request is shaped
The Video Router docs give the pattern. You send the model id, a prompt that states the change (their example is replacing a bottle with an apple and keeping everything else the same), a video_url that points at the source clip, and optionally a resolution. The default resolution is 720p. You do not send aspect_ratio or duration, because the edit follows the source. video_url is the edit source, not a reference, so it cannot be combined with image_url, end_image_url or the reference_*_urls fields.
Cost of a loop
Sume bills the provider list times 1.25 for each output second, as a function of resolution. The table assumes each pass outputs 8 seconds, a stand-in for your real source length. The poll response's usage.cost is the number to trust; use these rows to plan.
| Passes | Output seconds | At 720p ($0.125/s) | At 1080p ($0.1875/s) |
|---|---|---|---|
| 1 | 8 s | $1.00 | $1.50 |
| 2 | 16 s | $2.00 | $3.00 |
| 3 | 24 s | $3.00 | $4.50 |
| 5 | 40 s | $5.00 | $7.50 |
Keeping the loop cheap
Chain passes on the previous result rather than the original, because each instruction then builds on what you accepted. Trim the source to the shortest useful clip first; the edit costs output seconds, so a 4-second source is half the price of an 8-second one. Review at 720p and render the last pass at 1080p only when you are done.
Keep each prompt to one change. A prompt that asks for three changes and gets one wrong costs a whole pass to repair, while three one-change prompts let you stop when the clip is right.
A three-pass example
Imagine an 8-second product shot. Pass one: replace the bottle with an apple and keep everything else. Pass two: make the background a warm kitchen. Pass three: remove the shadow on the left. Each is a separate request with the previous result as the video_url, each at 720p, so each is about $1.00 and the whole loop is about $3.00. If you submit all three with an Idempotency-Key per pass, a retry after a network error returns the original job instead of charging a second time.
What the edit mode does not do
Native audio is always on for this model, and the API rejects generate_audio: false. Google's Omni page notes that voice editing is unsupported, so an edit loop is for picture changes, not for rewriting speech. Reference clips are limited to 3 seconds each, which is a separate limit from the edit source. If your source is longer than 10 seconds, trim it first with Sume's video trim, then edit the part that needs it.
Sources
Related posts
More in Media tools
- Group TTS sentences into 5 to 14.8 second lip-sync segments in Python
A short Python planner that groups sentence durations into segments the Sume H3 Max lip-sync route accepts (5 to 14.8 s) and prints the 768p cost of each.
- H3 Max lip sync: 480p drafts then a 1080p final for a 6-second line
A 6-second line costs $0.375 at 480p and $1.20 at 1080p on Sume H3 Max lip sync. Drafting pays only past 1.45 expected attempts. The math, with code.
- Half-banner video: a still stacked over a clip for $0.02
Sume timeline compose stacks one still and one video in a single frame for a flat $0.02 per job. Layout ratio, overlay mode, 300 s ceiling and a ready request.
- Holiday music bed under a voice: soundtrack duck_db 0 to 20
Sume Timeline takes a soundtrack bed with gain_db, loop, fade_out_seconds up to 10 and duck_db 0 to 20 under a voice spine. Python plan call and errors.
Written by Sume