gemini-omni-flash-preview shuts down on September 30. If you call it in production, your app throws errors on October 1. The migration is usually a one-line model ID change — but two silent breaking changes will bite you if you just swap the string and ship. Here is what to fix before the deadline.
Two Breaking Changes That Aren’t in the Headline #
Google’s API changelog lists the deprecation date and points you at gemini-omni-1.1-flash. However, what it doesn’t emphasize is that the output contract changed in two ways that will silently break existing code.
Breaking change one: large videos no longer return inline. In the preview, all output came back as inline base64 in output_video.data. In the GA release, videos above 4MB return as a URI instead. If your code reads output_video.data directly at 720p or above, it will silently return nothing. Add a URI check before you read the data field.
Breaking change two: resolution is now an explicit parameter. 720p remains the default, but the API now accepts 360p, 1080p, and 4K. If any part of your pipeline assumes a fixed output dimension or file size, verify your assumptions hold after migration. The 1080p and 4K options are upscaled — not natively generated at those resolutions — worth knowing before you promise clients “4K video generation.”
response = client.generate_video(
model="gemini-omni-flash-preview",
prompt="..."
)
video = response.output_video.data # silently fails at >4MB in 1.1
response = client.generate_video(
model="gemini-omni-1.1-flash",
prompt="..."
)
video = response.output_video.uri or response.output_video.data
What You Actually Gain With 1.1 #
The preview was not a great model to begin with. Its 1-second extension context window caused visible drift in long sequences: camera motion would reset, lighting would shift, characters would subtly change between clips. The GA release fixes this with a 10-second context window. According to Google’s Gemini Omni 1.1 Flash migration guide, sequences up to 40 seconds now hold together coherently — that’s the single most important production improvement.
First-and-last-frame interpolation is new and genuinely useful. Supply two images and the model generates a transition clip between them. Product reveals, camera orbits, seamless loops — these are use cases that simply didn’t exist in the preview.
Moreover, the 360p drafting mode is a cost-management tool worth building into your workflow immediately. It runs roughly 60% faster and costs about one-third of 720p generation. The right pattern: iterate at 360p until the prompt produces what you want, then render the final clip at your delivery resolution. If you’ve been burning API credits on 720p iteration, stop.
Pricing and Regional Limits #
Video output costs $17.50 per million tokens — approximately $0.10 per second at 720p. A 10-second clip runs about $1.00; the 40-second maximum via chained extension calls reaches about $4.00. The 360p drafting tier brings iteration cost down to roughly $0.033 per second. There is no free API tier — consumer access via Google Flow or the Gemini app is a separate entitlement that doesn’t carry over automatically.
One regional constraint worth noting: if you operate in the UK, Switzerland, or the EEA, you cannot edit or extend uploaded videos through the API. Text-to-video generation still works in those regions. This is a licensing constraint, not a technical one, and it applies to the GA release just as it did to the preview.
The Bigger Picture: Sora Is Gone, Omni Fills the Gap #
OpenAI shut down the Sora API on September 24. Google’s timing here isn’t subtle — Gemini Omni 1.1 Flash is now the default video generation API for developers who don’t want to operate their own models or switch to ByteDance’s Seedance 2.0. Furthermore, this deprecation is Google clearing the preview-era slate and committing to Omni 1.1 Flash as the stable production foundation — especially relevant now that Gemini has been accelerating its GA releases across all modalities.
September 30 is also just the first deadline in a broader cleanup cycle. According to Google’s full deprecation schedule, October 2 brings the retirement of gemini-2.5-flash-image, and October 20 takes out gemini-2.5-flash, gemini-2.5-flash-lite, and gemini-2.5-pro on Vertex AI. If you’re running a Gemini-heavy stack, your migration calendar just got crowded.
What to Do Today #
Three things. Change the model ID to gemini-omni-1.1-flash. Update your output-handling code to check for a URI before reading inline data. Test at the resolution your production workflow uses — don’t assume 720p behavior is identical to the preview. That’s it. Do it before September 30, not on October 1.