Reuse a Sora 2 prompt on Gemini Omni Flash: what to rewrite

Most of a Sora 2 prompt carries over to Gemini Omni Flash 1.1: the scene, the camera framing, the action beats and the lighting words. Four things need rewriting: the dialogue block, any shot longer than 10 seconds, the habit of leaving shot count to the model, and the way reference images are addressed. OpenAI's own video guide now says the Sora 2 models and Videos API were shut down on September 24, 2026, so these prompts need a new home.

This post compares OpenAI's Sora 2 Prompting Guide with Google DeepMind's Omni prompt guide, both read on 2026-10-03, and then shows the same idea as a Sume Video Router request. Sume's catalog id for the model is gemini-omni-flash-1.1. It is one of several ids you can call; this post only covers moving a prompt, not whether Omni is the right replacement for your clips.

What carries over unchanged

OpenAI's guide tells you to describe a shot like a storyboard sketch: framing, angle, depth of field, action in beats, then light and palette. Google's guide asks for the same ingredients in different words: shot size (wide, medium, close-up), camera movement, style, lighting source and quality, location, and action. A prompt that already names a wide establishing shot at eye level, a slow push in, soft window light and a single clear action does not need to be rewritten for Omni.

Google also says you do not need to describe every small detail of a location or every frame of a complex action. If your Sora prompt was long because you were fighting for consistency, try cutting it by a third before you spend a second render.

What to rewrite

The table sets the Sora guide habit beside what Google's page says and what the Sume request looks like. Where a page is silent, the table says so rather than guessing.

The shot-count and length rewrite

Sora's guide says the model follows instructions more reliably in shorter clips and that stitched short clips often beat one long generation. That advice transfers directly, because Omni stops at 10 seconds. If your Sora prompt described a 16 or 20 second sequence, split it into beats of 10 seconds or less and run one request per beat, then join the clips on the Timeline 1.0 surface.

Because Omni will otherwise cut between a few shots, a prompt that was written as one long take needs the line "In a single continuous shot" near the top. Google's guide also lists "one continuous shot" or "oner" as accepted wording, and "static", "locked off" or "fixed" for a camera that does not move.

The same prompt as a Sume request

This request carries the camera, light and action words over and adds the single-shot line. The reference tag is only needed if you also send reference_image_urls; leave it out otherwise. The request returns an async job, so store the job id and poll it, as described in Jobs and results.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: sora-port-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "In a single continuous shot, medium close-up from slightly behind. A woman takes four steps to the window, pauses, and pulls the curtain in the last second. Soft window light with a warm lamp fill; amber, cream and walnut tones.",
    "resolution": "720p",
    "duration": 8,
    "aspect_ratio": "9:16",
    "mode": "async"
  }'

What does not port

Sora-specific objects do not have a one-to-one match in the Video Router. Sora characters, the extension endpoint and the edits endpoint were API features of OpenAI's product, and the Omni request shape is different: reference images go in reference_image_urls (up to 10), edits go through video_url, and there is no reference_audio_urls. The Video Router docs list each capability and what it accepts. Read capabilities from GET /v1/video-router/models instead of assuming one envelope, and rerun your five most important Sora prompts as a small test before moving the rest.