Tutorial b-roll with Omni: six 10-second inserts for $7.50

A minute of b-roll for a tutorial video is six 10-second Omni inserts, which costs $7.50 at 720p on Sume ($0.125 per second), $11.25 at 1080p and $22.50 at 4K. Omni's range is 3 to 10 seconds, so 10 seconds is the longest single insert. Veo 3.1 only offers 4, 6 or 8 seconds, which means eight inserts of 8 seconds for 64 seconds. Read on 2026-10-08.

B-roll is a good fit for AI video because nobody checks it closely. It still has to match what the voice says, so write the shot list from the script, not from the tool.

Write the shot list from the script

Take each sentence of the voiceover that needs a visual and write one insert for it. Keep inserts to a single action. For a tutorial about brewing coffee, the line pour in slow circles gets one pour shot, not a whole barista scene.

• One insert per spoken idea; length equals the sentence length rounded up, within 3 to 10 seconds.
• Same aspect ratio as the main video: 16:9 or 9:16, since Omni offers nothing else.
• If the same hands, mug or desk appear in every shot, send a reference image for them; Omni on Sume takes up to 10.
• Say No dialogue so the inserts do not talk over your narration.

Cost of one minute of b-roll

The Veo rows are Google's direct prices and assume 8-second clips; they do not include a join.

Join and check

Inspect each clip's real length before you place it; Sume's video inspect and video trim steps exist for that, and trim returns a new MP4 for a [start, end) range. Omni clips carry their own sound, so mute the inserts or lower them under the narration. A per-minute figure ignores re-rolls: a quarter of your inserts will need a second try, so a budget of $9.40 at 720p is more honest than $7.50.

Keeping the look consistent

Inserts from different requests can look unlike each other. Reuse the same style sentence in every prompt (lens, light, color) and a reference image of your set or props. Sume takes up to 10 reference images per Omni request.

Submit with an Idempotency-Key header so a retry after a dropped connection returns the original job instead of a second charge; the same key with a different payload returns 409. Then poll GET /v1/jobs/:id/status and fetch GET /v1/jobs/:id/result when it finishes.