Gemini Omni 1.2 Flash Is on Vercel's AI Gateway, and It Generates Video

Google's Gemini Omni 1.2 Flash is now available on Vercel's AI Gateway, and it brings text-to-video generation into the same unified setup designers already use for language models. No separate Google account, no second API key, no extra dashboard to juggle.

That framing matters. Video generation has existed for a while, but it has usually lived behind its own login, its own billing, and its own SDK. Dropping it into AI Gateway changes the friction equation considerably.

What Gemini Omni 1.2 Flash Actually Does

The model generates short videos, between three and fifteen seconds, from a text prompt. It also works the other way: give it an existing clip (again, three to fifteen seconds) plus up to six reference images, and it will edit that footage. Output options cover 360p all the way up to 4K, in either 16:9 landscape or 9:16 portrait, so you can target a widescreen web layout or a vertical mobile feed without re-generating.

One feature worth paying attention to is the extension call. You can lengthen a generated clip by three to ten seconds per request, up to thirty seconds total. And if you want to refine a clip rather than start over, you pass a previousInteractionId parameter, essentially a reference to your last generation, and the model continues from there. That iterative loop is closer to how designers actually work: make something, react to it, adjust, repeat.

Native audio is included in generation, which is notable. Most early video tools treated audio as a separate pipeline, or left it out entirely.

Why the Gateway Part Matters as Much as the Model

Vercel's AI Gateway is a single layer that sits in front of multiple AI models. You call one unified API, it handles routing, usage tracking, cost reporting, retries if a provider goes down, and failover to a backup model. For someone building a vibe coded project without a dedicated infrastructure team, that is a meaningful safety net. If Google's endpoint has a bad moment, your app does not necessarily fall over.

The practical upshot: you are not adding a new, separate video API to your project. You are adding a model to a gateway you may already use for text. The billing, the logs, and the error handling all live in the same place.

How a Designer Could Actually Use This

The most direct use case is prototype-quality motion content inside an app. Think of an onboarding screen that shows a three-second animated demo of your product, generated on the fly from a prompt, rather than a static illustration you had to export from Figma. Or a travel app that generates a short landscape clip based on a destination the user searches for.

Here is how the workflow looks in practice, without getting into code. You pick google/gemini-omni-1.2-flash as your model inside AI SDK, write a prompt describing the scene, specify the resolution (360p is fine for quick prototyping) and duration, and the response comes back as a video file you can drop straight into your interface. The streamText path lets you handle that file as it arrives rather than waiting for the full response, which feels snappier in a live app.

For editing, the six-reference-image input is interesting for designers specifically. You could feed it screenshots or moodboard frames and ask it to generate motion that fits a visual language you have already established. That is not guaranteed to be precise, but as a starting point for exploring animated UI concepts it is genuinely useful.

The iterative refinement via previousInteractionId means you do not burn a fresh generation every time you want a small change. Adjust the prompt, reference the last clip, get a revision. That loop should feel familiar if you have used image generation tools.

What to Keep in Mind

A few honest caveats. The thirty-second total length is a hard ceiling for now, so this is not a tool for anything that needs sustained video. The model is described as "Flash" which in Google's naming convention generally signals speed and cost efficiency over maximum quality, so set expectations accordingly for high-fidelity output.

It is also worth noting that video generation is computationally heavier than text, so costs and latency will be higher than what you are used to from language model calls. Vercel's usage tracking inside the gateway will help you keep an eye on that, but it is worth building in some awareness before you wire this up to a feature that calls it on every user action.

The iterative refinement feature is promising but still a relatively new pattern. How well the model actually honors a previousInteractionId across varied prompts is something you will want to test rather than assume.

Overall, this is a meaningful expansion of what a designer building with AI can reach for without leaving their existing stack. Short, prompt-driven video is now one API call away from everything else.