Omni edit: one change per turn or one combined prompt, and the cost

A combined prompt is the cheapest way to edit a clip with Gemini Omni Flash 1.1 on Sume, at $1.25 for a 10-second clip at 720p, and three one-change turns cost three times that, $3.75. Choose turns when a failed edit would waste more than the extra $2.50.

The Video Router bills each video_to_video call on the output seconds, so the number of calls, not the number of changes, sets the price. This post prices both routes and gives a rule for picking.

The two routes

Combined: one call, a prompt like "Replace the bottle with an apple, warm the light, and change the table to oak. Keep everything else the same." Chained: three calls, one change each, the output of each fed as video_url to the next. The mechanics are on the Video Router page: video_url is the edit source, resolution is optional and defaults to 720p, and aspect_ratio and duration are not sent.

A rule for choosing

• Use one combined prompt when the changes are independent and small, such as a color plus a prop.
• Use turns when one change is risky, for example replacing a person or a prop that is in hand, and you do not want it to damage the others.
• If the first combined try fails, split it: you have paid $1.25 and learned which change was fragile.
• Do the fragile change first. A later change builds on a clip you have already approved.

Expected cost with a failure rate

Assume a combined edit works 50 percent of the time and each single-change edit works 80 percent of the time. The numbers are placeholders, not Sume data. Expected combined cost is $1.25 divided by 0.5, or $2.50. Expected chained cost is three turns each divided by 0.8: 3 x $1.25 / 0.8 = $4.69. With these odds the combined route is still cheaper; the chain wins only if its per-turn success rate is much higher than the combined one, or if a failed combined attempt ruins a clip you cannot regenerate.

Measure your own rates over twenty clips. Two columns in a sheet, attempts and accepted, are enough to replace the placeholders.

Keep resolution out of the argument

Iterating at 720p keeps each turn at $1.25 for ten seconds. Moving the last turn to 1080p costs $1.88 instead. A 4K last turn is $3.75, which makes it the one place where the choice of turn count is dominated by the choice of final resolution, so run the experiments at 720p first.

Wording that holds a combined edit together

Combined prompts fail when they read as a list of unrelated wishes. Give the model one scene description rather than three commands: "The same kitchen scene, but the bottle is an apple, the light is warm morning sun, and the table is oak. Keep the camera move and the people unchanged." Name what stays, as the Sume docs example does with "Keep everything else the same."

For a turn-by-turn chain, the same clause is even more important. A turn that does not say what to keep gives the model room to change it.

A worked batch

Take 20 clips of 8 seconds at 720p. Combined edits cost $1.00 each, $20.00. Three-turn chains cost $3.00 each, $60.00. If the combined route needs a second try on a third of the clips, add $6.67 and you are still at $26.67. The chain's extra $33 buys per-step approval; whether that is worth it depends on how costly a bad clip is when it reaches a client.

Sound comes with the edit

The model always produces native synced audio and the API rejects generate_audio: false, so an edit changes a clip that has sound. If your first prompt is about the picture only, say "keep the audio the same" and listen to the output before you accept it.