Gemini Omni Flash: what shipped and what Google held back
Google launched Gemini Omni Flash at I/O 2026 — here's what the first Omni model does, what was deliberately withheld, and the practical move for creators on OmniArt.
Google I/O 2026 landed on May 19, and by the time the keynote ended Gemini Omni Flash was live. Not "coming soon," not "limited preview" — live, the same day. Two weeks ago we published our read of the pre-I/O leaks, separating the confirmed signals from the speculation. Now we have the actual model. Here is what shipped, what Google deliberately did not ship, and what it means for creators who have work due this week.
Omni Flash is the first public model in Google's new "Omni" framework. It is not Veo 4, and it is not a rebrand of Veo 3.1 — it is a separate product line, with a higher-tier Omni Pro already confirmed by Google DeepMind as a follow-up. No date on Omni Pro. Flash is phase one.
What confirmed vs. what was withheld
The leak piece called the model "Gemini-native video with omni-modal ambitions." That held up well. Here is the full picture now that the keynote dust has settled.
| Feature | Status | What it means for creators |
|---|---|---|
| 10-second video clips with synchronized audio from a single prompt | Shipped | Short-form social, trailers, and idents are the natural fit at this clip length |
| Any-to-any input: text, image, audio, and video in one prompt | Shipped | You can hand it a reference image, a voice note, and a brief — one prompt grammar for all three |
| Conversational / chat-based editing ("change the lighting", "swap the dog for a cat") | Shipped | The workflow shift the leak piece flagged as the real headline — more on this below |
| SynthID watermark in every output | Shipped — non-optional, no API toggle | Plan for watermarked output by default; check use-case terms before commercial placement |
| Editing speech or audio inside generated video | Held back for safety | Deepfake-adjacent risk; Google has confirmed it is withheld deliberately, not a capability gap |
| Avatar mode | Held back | Same category of concern as audio editing — no timeline given |
| Developer API | "Coming weeks" | Do not build a production pipeline against it until the API is live and stable |
Warning
Google has also publicly acknowledged three current limitations: visual consistency during edits, complex motion sequences, and rendering readable in-video text. These are the same weak spots the broader AI video category shares; Omni Flash has not solved them.
The leak vs. reality check
Before I/O we outlined three scenarios for what Omni could be: a consumer rebrand of Veo, a Gemini-native video model, or a true omni-modal unified system. We called "a blend of scenarios 2 and 3" as most likely.
That was accurate. Omni Flash is unmistakably Gemini-native — it runs inside the Gemini app and Google Flow, not as a standalone Veo surface — and it is genuinely any-to-any in its inputs. The "omni-modal" framing Google has given it is not marketing overreach; combining text, image, audio, and video into one prompt is a real capability change from Veo 3.1's input model.
What the speculation got wrong: the leaked "remix" framing undersold the depth of the conversational editing feature. It is not just remix-from-scratch. It preserves consistency across multi-turn edits, which is a materially different thing.
Conversational editing is the real headline
Every major AI video model today works the same way at the workflow level: you write a prompt, wait, download the clip, and re-prompt if it is wrong. Omni Flash breaks this. The conversational editing feature lets you type "change the lighting to golden hour" or "swap the dog for a cat" and get a revised clip that maintains consistency with the prior outputs rather than regenerating from scratch.
This matters because the cost of iteration in video has always been the regenerate cycle — both in time and credits. Multi-turn editing that preserves consistency compresses the gap between a first draft and a finished clip. It also means the model holds state about your project in a way that generate-and-discard workflows don't.
The current acknowledged limits are real: complex motion sequences lose coherence across edits, and the model can still drift on fine-grained visual details. But the workflow principle is sound, and it is the feature most likely to age well as the underlying model improves.
Where Omni Flash fits in the lineup
Omni Flash's strengths are consumer accessibility, conversational iteration, and multi-modal input flexibility. Its limits — 10-second clips, no speech editing, acknowledged motion and text rendering gaps — define its lane clearly.
| The shot needs | Reach for |
|---|---|
| Conversational iteration, chat-based refinement | Omni Flash (on Google's surfaces) |
| Native 4K, spatial audio, broadcast finish | Veo 3.1 |
| Long single takes | Sora 2 |
| Multi-shot storyboard continuity | Kling, V6 + BACH |
| Fast, stylized, high-energy clips | PixVerse models |
| Value at volume | Kling for cost-efficient finished seconds |
For a deeper look at how Omni Flash and Veo 3.1 compare shot-for-shot, see Gemini Omni Flash vs. Veo 3.1: which one for your workflow.
Where to actually access it
Omni Flash is live on YouTube Shorts, YouTube Create, the Gemini app, and Google Flow. Pricing runs through Google's AI subscription tier: AI Plus starts around $7.99/month, and Ultra dropped from $250 to $100/month. A developer API is arriving "in the coming weeks" — no exact date.
For context on the broader Veo line, Veo 4 release status and where Veo fits on OmniArt covers what Veo 3.1 already does and how it sits inside a multi-model workspace.
Omni Pro is confirmed — but unscheduled
Google DeepMind has confirmed a higher-tier Omni Pro is coming, described as "a step change above Flash." There is no release date, no feature list, and no preview access. Plan around what ships, not what is promised.
If your pipeline has a Q3 deliverable, build it against Omni Flash's confirmed specs today. When Omni Pro lands, you add it as an option inside a workflow that is already producing — you do not wait for it, and you do not re-platform for it.
Note
What to do this week
Omni Flash lives on Google's own surfaces — the Gemini app, YouTube Shorts, Google Flow. If you want to test conversational editing, that is where to do it. Google has not announced third-party API integrations beyond the "coming weeks" developer timeline.
On OmniArt, you ship today with Veo 3.1 for native 4K and spatial audio, and the rest of the lineup — PixVerse models, Sora 2, Kling, HappyHorse, Seedance 2, and more — across image, video, audio, and music in one workspace. One balance, one prompt grammar, one place to compare outputs side by side.
For the practical steps on getting the most out of Veo 3.1 while you evaluate Omni Flash, the Veo 3.1 prompt and cinematic guide covers the full workflow from brief to finished clip.
The practical move: run your current brief through the models that are live and stable. When Omni Pro lands — or when the Omni Flash API opens — you add it to a pipeline that is already producing, rather than waiting to start.
FAQ
Is Gemini Omni Flash available right now?
Yes. It launched at Google I/O 2026 on May 19, 2026, and went live the same day via YouTube Shorts, YouTube Create, the Gemini app, and Google Flow. A developer API is described as "coming weeks."
What is the difference between Omni Flash and Veo 3.1?
Omni Flash is Gemini-native, accepts any-to-any inputs (text, image, audio, video in one prompt), and has conversational multi-turn editing. Veo 3.1 is a dedicated video model with confirmed native 4K output and spatial audio. They have different strengths and currently live on different surfaces.
What features did Google hold back from Omni Flash?
Two capabilities were deliberately withheld: in-video speech and audio editing, and avatar mode. Google has confirmed these were held for safety reasons, not because of technical limitations. There is no release date for either.
Will Gemini Omni Pro replace Flash?
Google DeepMind has confirmed Omni Pro as a future higher-tier model described as "a step change above Flash," but no features, pricing, or release date have been disclosed. Plan around Flash's confirmed capabilities; treat Omni Pro as a future addition.
Does Omni Flash have a SynthID watermark?
Yes. Every Omni Flash output includes a SynthID watermark. It is non-optional and has no API toggle. Check the platform's terms of service before using outputs in commercial contexts.
Ready to Create?
Start generating amazing content with AI