Gemini Omni Flash is Google’s low-latency, conversational video model: a system you talk to in plain language to generate and edit short video clips, then refine them turn by turn without starting over. Google introduced it at its I/O developer conference on May 19, 2026, and describes it as the first model in a new "Omni" family, an any-to-any system that accepts text, images, audio, and video as input and returns video as output. The "Flash" in the name follows Google’s convention for its fast, lower-cost model tier, so the pitch here is speed and interactivity rather than maximum cinematic polish.
The short version: most AI video tools work like a vending machine, where you write a prompt, wait, and get a finished clip you either accept or throw away. Gemini Omni Flash is built to work more like a conversation. You describe a shot, watch what comes back, then say what to change, and the model holds the context of the exchange so you are not re-explaining the scene every time. That shift, from one-shot generation to iterative editing, is the whole point of the product. This piece covers what it is, what it can and cannot do, where you actually use it, how its provenance layer works, and how it stacks up against dedicated video models like Google’s own Veo.
What "conversational video" means
Conventional text-to-video generation is a single transaction. You submit a prompt, the model renders a clip, and any change means writing a new prompt and generating again from scratch, often losing the character, framing, or style you liked in the first attempt. Google’s framing for Gemini Omni Flash is that video should behave more like a chat thread.
According to Google, the model maintains conversational context across multiple turns, so you can ask for a clip, then say "make it dusk," then "now have the character turn toward the camera," and it applies each change on top of the last rather than treating every request as a fresh start. Google also says the model keeps characters consistent across edits, which is one of the harder problems in AI video and the thing that usually breaks when you regenerate.
The other half of the framing is world knowledge. Google positions Omni Flash as combining a generative video engine with Gemini’s broader reasoning, so it is meant to reason about what should plausibly happen next in a scene, drawing on an intuitive sense of physics plus Gemini’s knowledge of history, science, and culture. That is a marketing claim as much as a technical one, and worth testing against your own footage, but it explains why Google filed the model under "Omni" rather than simply shipping another text-to-video endpoint.
What Gemini Omni Flash can do
Per Google’s model card, Gemini Omni Flash generates and conversationally edits up to 10-second, 720p video from text, image, and video inputs. A few capabilities stand out:
- Any-to-any input. Text, images, audio, and video can all go in, and video comes out. You can start from a written description, seed it with a reference image, or feed it an existing clip to modify.
- Multi-turn editing. The model refines a clip through natural language across several turns, holding context so each instruction builds on the last instead of regenerating from zero.
- Character consistency. Google says characters stay consistent across edits, which is what makes iterative refinement usable rather than a fresh roll of the dice each time.
- Native provenance. Every output carries a SynthID watermark and ships with C2PA provenance credentials in a standard MP4 file, covered below.
The deliberate limit is duration and resolution. Clips top out at 10 seconds and 720p. Google frames the 10-second cap as a deployment choice rather than a hard model constraint, which suggests longer outputs may follow, but as shipped it is a short-clip, standard-definition tool, not a 4K cinema renderer.
Where you use Gemini Omni Flash
There are three practical entry points, aimed at three different users.
For most people, the model shows up inside the Gemini app. Google says it is rolling Omni Flash out to Google AI Plus, Pro, and Ultra subscribers globally, so if you have a paid Gemini plan, conversational video generation appears as part of the assistant you already use. If you are new to Google’s consumer AI surfaces, our Gemini Spark explainer covers how that side is organized.
For creators, it also lands in Google Flow, Google’s AI filmmaking tool, where it fits alongside timeline and shot-based workflows for assembling generated clips into something longer.
For developers, Omni Flash is available through the Gemini API, so you can call it from your own application the way you would any other Gemini model. If you are prototyping, Google’s browser-based AI Studio is the usual place to test prompts before wiring the API into production code.
The provenance layer: SynthID and C2PA
One feature is not optional: every clip Gemini Omni Flash produces is marked as AI-generated. Google embeds a SynthID watermark, its imperceptible, machine-detectable signal, into each output, and the clips also ship with C2PA content-credential metadata inside a standard MP4 file. SynthID has been Google DeepMind’s provenance tool since 2023, and C2PA is the cross-industry standard for tamper-evident "content credentials."
For a business, this matters in two directions. It means anything you generate is traceable as synthetic, which is useful for disclosure and internal governance, and it means the label is difficult to strip. If your organization has policies about labeling AI-generated media, this is a model that labels by default rather than leaving it to you. It is the same provenance thinking Google applies to its image models, including Imagen and its consumer image tools.
How it compares with Veo and other video models
The most important comparison is inside Google’s own lineup. Veo is Google’s dedicated, cinema-oriented video model, optimized for single-shot quality, higher resolution, and longer scene extension. Gemini Omni Flash is the opposite trade: a multimodal system that happens to output video, optimized for speed, interactivity, and cheap iteration rather than maximum fidelity. If you want one polished hero shot, Veo is the tool; if you want to talk your way to a rough clip and refine it quickly, Omni Flash is.
On price, Google lists Omni Flash at roughly $0.10 per second of 720p output, so a full 10-second clip runs about a dollar, with a batch discount roughly halving that. That undercuts Veo’s higher-resolution tiers, which is consistent with the "Flash" positioning as the fast, low-cost option.
Against the wider field, the tradeoff holds. Rivals like ByteDance’s Seedance line push single-clip length and resolution much further, generating longer, higher-resolution shots in one pass. Independent testers generally place Omni Flash’s raw generation quality below those dedicated models. Its edge is not the prettiest frame; it is the conversation, the multimodal inputs, and the built-in provenance. Choose on what you actually need: fidelity and length, or speed and iterative control.
Limits and what to check first
A few things to keep in mind before you build on it:
- It is new. Availability, pricing, plan limits, and quotas can change quickly on a just-launched model. Confirm the current figures for your region and plan in Google’s documentation before committing.
- Short and standard-def, for now. Ten seconds at 720p is the ceiling as shipped. If you need longer or higher-resolution output, that is a different tool today.
- Quality is not the headline. If pixel-level polish is your priority, a dedicated model will likely beat it; Omni Flash trades some fidelity for speed and editability.
- Everything is watermarked. That is useful for disclosure, but plan for it: outputs are labeled as AI-generated by design.
Understood that way, Gemini Omni Flash is not Google’s answer to the highest-quality video generator. It is an attempt to make video generation feel like a conversation, with provenance baked in from the first frame.
Frequently Asked Questions
What is Gemini Omni Flash?
Gemini Omni Flash is Google’s low-latency, conversational video model, introduced at Google I/O on May 19, 2026. It is the first model in a new “Omni” family, an any-to-any system that accepts text, images, audio, and video as input and returns video. It generates and conversationally edits up to 10-second, 720p clips, keeps characters consistent across edits, and watermarks every output.
What does “conversational video” mean?
Instead of writing one prompt and getting a single finished clip, you generate a clip and then refine it through natural language across multiple turns. The model holds the context of the exchange, so each instruction (“make it dusk,” “turn the character toward the camera”) builds on the last rather than regenerating from scratch. That iterative, chat-style editing is the model’s defining feature.
How do you access Gemini Omni Flash?
Three ways. Google is rolling it out to Google AI Plus, Pro, and Ultra subscribers inside the Gemini app; it also appears in Google Flow, Google’s AI filmmaking tool; and developers can call it through the Gemini API. AI Studio is the usual place to test prompts before integrating the API.
How much does Gemini Omni Flash cost?
Google lists it at roughly $0.10 per second of 720p output, so a full 10-second clip costs about a dollar, with a batch-processing discount roughly halving that. Because it is newly launched, confirm current pricing, plan limits, and quotas in Google’s own documentation before you rely on any figure.
How is it different from Google Veo?
Veo is Google’s dedicated, cinema-oriented video model, built for single-shot quality, higher resolution, and longer scene extension. Omni Flash is a multimodal system that happens to output video, built for speed, conversational editing, and cheap iteration rather than maximum fidelity. Use Veo for one polished hero shot, and Omni Flash to talk your way to a rough clip and refine it fast.
What are the limits of Gemini Omni Flash?
As shipped, clips top out at 10 seconds and 720p resolution. Google describes the 10-second cap as a deployment choice rather than a hard model constraint, so it may change. Independent testers generally rate its raw generation quality below dedicated video models. It is a fast, iterative, short-clip tool, not a 4K cinema renderer.
Does every clip get watermarked?
Yes. Every Gemini Omni Flash output carries an imperceptible SynthID watermark and ships with C2PA content-credential metadata inside a standard MP4 file. The labeling is applied by default and is designed to be difficult to strip, which helps with disclosure and internal governance around AI-generated media.
Who is Gemini Omni Flash for?
It suits marketers, creators, and developers who want to produce and iterate on short video clips quickly and cheaply, and who value conversational editing and built-in provenance over top-end fidelity. If your priority is the longest or highest-resolution single shot, a dedicated model will fit better. Evaluate it on your own material before committing.
