Gemini Omni video references: 3 clips of 3 s each, $1.31 all in

Gemini Omni Flash 1.1 on Sume accepts up to 3 reference videos in reference_video_urls, and each must be 3 seconds or shorter. If your source clips are longer, trim them first. Three trims at $0.02 each, plus a 10-second 720p Omni render at $1.25, come to $1.31.

Cost table

All prices are Sume billable prices.

Trimming to 3 seconds

Video trim is a flat $0.02 per job, with a source up to 1800 s and an output up to 900 s. precision: "exact" is the default and re-encodes frame-accurately; keyframe copies the stream but the cut can start a GOP early. For a reference clip that must start on a precise frame, leave the default. Send duration: 3 with a start.

{
  "video_url": "https://media.sume.com/artifacts/artf_demo/long.mp4",
  "start": 12.0,
  "duration": 3,
  "precision": "exact",
  "audio": "drop"
}

Limits to remember

Omni's text-to-video range is 3 to 10 seconds, 16:9 or 9:16, and 360p to 4K. The reference mode also takes up to 10 reference images. Refer to the media in the prompt as <IMAGE_REF_0> and <VIDEO_REF_0>, counted from zero in list order.

A 10-second 720p clip is $1.25 here, so the three trims add about 5 percent (0.06 / 1.25 = 4.8).

When the references are already short

If each reference clip is already 3 seconds or shorter, skip the trim and the job is just the render price, $1.25 for 10 seconds at 720p. Trim only the clips that are too long, and use the Video inspect probe to read each duration first. The trim route accepts a source up to 1800 seconds, so a long source can be trimmed down directly.

Cutting audio from the references with audio: "drop" is optional and changes only the reference file, not the Omni render price.

Why three seconds

The Video Router limit is a router rule: reference_video_urls takes at most 3 items and each item at most 3 seconds. A longer clip is not clipped for you, so send a trimmed reference or the request is refused. Trimming also lets you choose the moment, such as the three seconds with the motion you want the model to copy, instead of the first three seconds.

The prompt refers to each reference by position: <VIDEO_REF_0> is the first in the list, <VIDEO_REF_1> the second. Keep the list order the same as the order in your prompt text.

Planning a batch

For a batch of 10 such clips the trims are 10 x 3 x $0.02 = $0.60 and the renders are 10 x $1.25 = $12.50, so $13.10 in all, which is 10 times the single-job $1.31. If the same three references serve every clip, trim them once and reuse the trimmed files: the batch is then 3 x $0.02 + 10 x $1.25 = $12.56.

The saving from reuse is $0.54, small next to the render cost, but it also keeps the references identical across the batch, which matters if you compare results.

Output constraints

Omni text-to-video renders 3 to 10 seconds at 360p up to 4K, in 16:9 or 9:16 only. For a TikTok vertical ad choose 9:16; the 720p output is well above the 540x960 minimum on TikTok's non-Spark page.