Nathan Smith · 27 October 2026

Google Genie 4: DeepMind's new AI world model is incredible

These are not videos. Google DeepMind's Genie 4 builds photorealistic worlds live, with sound and characters who talk back, and lets you walk around inside them. Here is what it does, when you can try it, and what it means for brands.

Every so often a demo lands that makes you stop scrolling and watch it twice. Google DeepMind's Genie 4 is one of those. It looks like a film, but nothing in it was filmed: it is a world being made up on the spot, one frame at a time, while somebody walks around inside it with a keyboard.

Here is what it is, what it can do, and why we think it matters for anyone who makes video, games or ads.

Look out for the crane walking across the snow to a piano, the painting you can wander into, and the scooter scene where the characters actually speak.

So what is a "world model"?

Most AI video tools you will have seen (Veo, Sora, Kling and friends) work like a director who shoots a clip and hands it to you. You write a prompt, wait, and get back a finished few seconds of video. You can watch it, but you cannot step into it.

A world model works more like a video game engine that has never been programmed. You describe a place, and it generates that place live, many times a second, reacting to what you do. Press W and you walk forward. Turn the camera and the model works out what should be behind you. Nothing is stored or pre-rendered: every frame is invented in the moment, based on everything that came before it.

That is the trick that makes the demo so unsettling. These are not videos. They are places you can move through.

What is new in Genie 4?

Genie 3 (August 2025) was the first of these models you could actually walk around in, at 720p and 24 frames a second. Impressive, but it looked like a slightly dreamy video game. Genie 4 takes a big step towards looking real. The headline changes:

Photorealism. DeepMind says Gemini Omni keeps each world anchored to how things really look and behave, while other Gemini models handle the reasoning that keeps a scene coherent from one moment to the next.
Sound, generated live. Every world now makes its own audio as you go: ambient noise, music, and dialogue. It is DeepMind's first real-time world model to do that natively.
People who react. Characters move naturally, respond to their surroundings, and can hold a conversation with you in real time.
Steer it with words. Mid-scene, you can type "make it snow", "a shooting star crosses the sky" or "add a person on the bridge", and the world changes around you without starting again.
Better physics. Things like collisions and aerodynamics behave more like you would expect.
Longer scenes. Google has not put a number on it, but early reports suggest scenes now hold together for around 2 minutes, roughly double what Genie 3 managed.

Can I try it?

Not quite yet, unless you are on the list. Genie 4 is a research preview with a small group of trusted testers. Google says it is coming soon to Google AI Ultra subscribers through Project Genie, and to early-access partners in the Gemini API and Google AI Studio. There is a waitlist for Project Genie if you want to be near the front of the queue.

Everything it makes, picture and sound, carries Google's SynthID watermark, so it can be identified as AI-generated later.

What is it actually for?

DeepMind talks about world models as a step towards more general AI, and one of the most practical uses is training robots and AI agents in simulated worlds before they are let loose in real ones. (The film shows a robot arm rehearsing a lab task.)

But the testers it has picked tell you where the nearer-term interest is: filmmakers using it for pre-visualisation, photojournalists turning old 2D photos into places you can step into, and game developers prototyping levels from a sentence. Bloomberg pointed out that the previous version was enough to rattle video game stocks, which gives you a sense of how seriously the industry is taking it.

What this means for brands and marketers

It is early, and nobody is going to be running a Genie 4 ad campaign next week. But a few things are worth thinking about now.

Pre-vis gets cheap. Being able to walk a client through a location, a set or a store layout before a single thing is built or booked changes how creative ideas get sold in.
Ads you can step into. Interactive product worlds (a car you can drive round a coastal road, a hotel you can wander through at sunset) stop being a big-budget 3D project and start being a prompt.
Video is becoming a place, not a file. For years the deliverable has been a clip. Models like this hint at a future where the deliverable is a world, and the clips are just moments you capture inside it.

We are already making AI films for clients with today's generation tools, and the jump from Genie 3 to Genie 4 is a good reminder of how quickly that toolkit is moving. If you want to see what is possible right now, have a look at our AI Video work.

The honest caveats

DeepMind is upfront that this is not finished. Detail can drift over longer sessions, especially after lots of edits; physics can still go wrong when things get complicated; on-screen text is hard to read; and the audio does not always line up. Those are the same gremlins every AI video tool has, just in a much more ambitious setting.

Still, watch the film again and remember that none of it was filmed. That is quite something!