Google's Gemini Omni turns video editing into a conversation

What Omni Flash actually ships with today

Google is releasing the first model in a new family — Gemini Omni Flash — and framing it as the successor to last year's Nano Banana image work. The pitch is a single model that takes images, audio, video and text as input and produces high-quality video as output.

Omni is our new model that can create anything from any input — starting with video.Montana Labs

The phrase "starting with video" is doing a lot here. Output is video-only at launch; Google says image and audio output will come "in time." So despite the "anything from any input" framing, Omni Flash is currently an any-input, one-output-modality system.

The conversational editing loop is the real product claim

The most concrete differentiator in the announcement is stateful, multi-turn editing. Google's example carousel walks through a violinist clip that is progressively transported to a new environment, has the violin made invisible, then shifts to an over-the-shoulder camera angle — each instruction building on the last.

Google describes the value directly: "Every instruction builds on the last. Your characters stay consistent, the physics hold up and the scene remembers what came before." That memory-across-turns behavior is what separates an editing tool from a one-shot generator, and it's the capability the demos are built to prove.

The company also leans on physics and world knowledge as a distinction, citing improved handling of gravity, kinetic energy and fluid dynamics, plus examples like a claymation protein-folding explainer and a 26-letter alphabet montage. The recurring argument is that Omni reasons about what should happen next rather than matching patterns.

Audio inputs and avatars ship narrow on purpose

The input story has explicit caveats. Of the four input types, audio is the most restricted: "only voice references will be supported for audio to start," with other audio input types promised later. Several showcased prompts — synchronizing harp sounds to touched fern leaves, beat-matched style shifts — depend on richer audio input that isn't fully available yet.

On likeness, Google is shipping Avatars so users can generate video in their own voice, but it explicitly holds back the broader capability of editing audio and speech in arbitrary video, saying it is "still working to test this and better understand how we can bring this capability to users responsibly." Every Omni video carries the imperceptible SynthID watermark, verifiable through the Gemini app, Gemini in Chrome and Google Search.

The distribution move: free on YouTube Shorts, paid in Flow

The rollout plan is the part worth reading closely. Omni Flash goes to paying Google AI Plus, Pro and Ultra subscribers through the Gemini app and Google Flow, but it also arrives "at no cost" on YouTube Shorts and the YouTube Create App this week.

Putting a video-generation model into Shorts for free is a demand-side bet: it seeds Omni output directly into one of the largest short-form video surfaces before developers and enterprise customers get API access, which Google says comes "in the coming weeks."

The specific implication: editing statefulness is the thing to test

For teams evaluating Omni, the claim that matters isn't photorealism — it's whether the scene genuinely "remembers what came before" across many edits without drifting on character identity, physics or continuity. That persistence, if it holds up beyond curated demos, is what would make Omni usable as an editing tool rather than a generator you re-prompt from scratch.

The honest read on today's launch: video-in/video-out with conversational refinement is live and broadly distributed, while richer audio inputs, non-video outputs, general speech editing and API access are staged for later. The gap between "create anything from any input" and what Omni Flash does this week is the roadmap Google is asking users to fill in over time.