Gemini Omni Flash 1.1 limits for a Sora remix: 10 images, 3 videos

gemini-omni-flash-1.1 on Sume takes 10 reference images, 3 videos of 3 seconds each, and a 20000 character prompt. Tokens like <IMAGE_REF_0> point at inputs.

4 min readSume
All posts

gemini-omni-flash-1.1 on Sume takes up to 10 reference images and up to 3 reference videos of at most 3 seconds each, with a prompt of up to 20000 characters. It accepts no audio references. Edit mode takes a video_url, optional resolution, and rejects aspect_ratio. These are the numbers to check before you port a Sora remix flow.

Limits and prices

gemini-omni-flash-1.1 limits (read 2026-10-07)
ItemLimit
Clip length3 to 10 seconds
Resolution360p, 720p, 1080p, 4K
Aspect ratio16:9 or 9:16
Reference imagesUp to 10
Reference videosUp to 3, each at most 3 seconds
Reference audioNot accepted
Prompt lengthUp to 20000 characters
Audio outputNative, always on
Billed price per second360p $0.0375, 720p $0.125, 1080p $0.1875, 4K $0.375

Where remix went

OpenAI's Sora guide, read 2026-10-07, shows a remix endpoint marked deprecated, and the API itself ended on September 24, 2026. A remix sent an earlier video plus a changed prompt. The Sume equivalent sends your own video as video_url in edit mode and describes the change in the prompt.

The reference limits are the part with no Sora precedent. Sora remix took one earlier video; Omni can take several images and short clips at once, which allows prompts that combine a character from one input with a motion from another.

Reference tokens

Inputs are addressed inside the prompt by index. Image 0 is <IMAGE_REF_0>, the first reference clip is <VIDEO_REF_0>, and so on. A prompt like: keep the camera move of <VIDEO_REF_0>, put the jacket from <IMAGE_REF_0> on the lead, uses two inputs without any extra field.

Edit mode rules

Edit mode differs from generation in three ways, each documented. Resolution is optional and defaults to 720p. aspect_ratio is rejected, since the output follows the source. And duration is only a reserve hint, so the final charge comes from the finished job's usage.cost.

Twenty thousand characters is generous. Keep prompts short anyway: a clear paragraph with the reference tokens in the places they apply beats a long list of instructions that contradict each other.

  • Send the source as video_url.
  • Skip aspect_ratio, or the call returns 400.
  • Read usage.cost from the poll response after completion.

Planning around the caps

A 10 second 720p generation is 10 x $0.125 = $1.25 and a 4K one is $3.75. The 3 second cap on reference clips is the limit that bites most: trim clips first, or cut the single moment you want copied.

Where next

The request fields are in the video docs. For the full catalog see the twelve-id list, and for a comparison with Runway's video-to-video see this post.

Sources

Related posts

More in Models

All Models posts

Written by Sume