Google Genie 4 Explained: Type a Sentence, Walk Inside the World (and How to Prompt It)

A crane walks across the snow. Nobody filmed it and nobody animated it. Someone typed six words, "a shooting star crosses the sky," and the sky obeyed while they were still walking around inside it.

That's Genie 4, from Google DeepMind, announced October 27, 2026. Like every genie, it grants wishes, and every wish comes with fine print. We went through every demo clip Google published, all 12 prompts behind them, and the 47-minute interview with the team that built it. Here are the four wishes, the catch under each one, and how to write a prompt the way Google's own team does.

What a world model is (Genie 1 to 4)

In plain words: a video model makes you a movie to watch. A world model puts you inside the movie. You press a key to walk forward, and it draws the next frame, live, around you.

And it's moving fast:

• Genie 1 made blurry 2D platform games.
• Genie 2 went 3D, but it wasn't real-time.
• Genie 3 made worlds you could walk through live, and the public got to try it this January.
• Genie 4 adds sound, people who react to you, and a second AI steering the whole thing.

Wish 1: change the world while you're standing in it

One demo starts as a single photo: a man playing saxophone on a floating island. The user types "a whale's tail rises behind him," and the water moves. Then they ask for snow.

The interesting part is a side panel. That's Gemini rewriting the wish so it fits the scene: keep the saxophone playing, keep the 16mm film look, keep the world consistent. You make a rough request, and Gemini turns it into a careful one.

The fine print: Google says it plainly. Visual details and object permanence (things staying where they were) can drift over time, especially after several prompt edits or as you move around. A viewer caught it in the crane clip: the shooting star goes behind the crane's head and never comes out the other side.

Wish 2: make it obey physics

A watermelon sits on a stone plinth. The prompt: drop a heavy white marble bust on top of it. The melon bursts, and pieces roll off the edge.

Then a single maple seed spinning down into the grass. That prompt is only 15 words: a macro close up of a maple seed falling from the tree into the grass. There's also a ribbon of pink silk the length of a river, rippling in a wind nobody animated.

The fine print: Google still lists complex physical interactions as unsolved. One drop, one fall, one breeze works. A long chain of things hitting each other is where it gets shaky.

Wish 3: people who react to you

Google calls this the most dramatic leap. In a busy Japanese train station in the 1980s, every passenger moves on their own. Side by side with Genie 3, the faces and the light are clearly better.

And now people react to you. In the team's live demo, an engineer walks through a cafe and drops a cup of coffee. A person in the world answers: "No worries at all, accidents happen." Nobody wrote that line. The world made it up on the spot.

The fine print: in the launch video, a dragon rider says "let's attack the castle." The reply: "No, it's a trap." And we never see the fight. Big, messy action is exactly what Google says it's still working on.

Wish 4: let me hear it

Take an old black and white jazz photo. Genie 3, the previous version, brings it to life with no sound at all. Same photo in Genie 4, and the band plays.

Google says this is its first real-time world model that makes its own sound. The music, the wind, the wingbeats of a dragon in flight: it all comes out of the same model as the pictures.

The fine print: real-time audio accuracy is still on Google's own list of open problems. We'll come back to why.

How Gemini directs the world, and why robots need it

Genie doesn't work alone. Genie draws the world. A second AI, Gemini 3.8 Flash, watches every frame like a film director. It asks three questions, over and over:

1. What just happened?
2. What did the user want?
3. What should happen next?

It splits your goal into steps, checks that each step really happened, and if you change the world, it rewrites the plan.

This is where it gets serious. Raia Hadsell, who leads this research at Google DeepMind, says we all carry a world model in our heads, one that predicts a little bit forward. It's how you catch a cup before it falls. Robots don't have that yet.

So in one demo, a bigger Gemini model called Argon plans a task step by step, and Genie renders a robot doing it. A humanoid plays checkers. A pair of robot hands lays out letter tiles. A lab arm pipettes liquid. No real robot, and nothing real gets broken.

Why not just use a game engine?

A viewer under the team's interview asked the smart question: an AI model can already write a whole 3D game in code, so why do we need Genie? Raia asked herself the same thing.

Here's the trade. Code remembers: walk away, come back, and the house is still there. Pixels improvise: a dragon flies in, or the whole world turns to stained glass. Her bet is a hybrid. Code keeps the house standing, and Genie makes the world come alive.

Steal the prompt: what Google's 12 prompts have in common

We pulled all 12 prompts Google published and counted. They add up to 1,828 words, and they follow a pattern.

The rules read like a director's notes. No CGI look. No video game look. No camera floating outside the body. Even the 434-word prompt stops to ban the fake look.

The team's own test prompts are telling too. Diego Rivas, from the Genie team, goes to a busy restaurant kitchen: lots of people moving at once, which is the hardest test there is. Raia goes the other way: bumper cars where every car is a different animal. One tests realism. The other tests imagination.

And the kitchen demo shows the audio fine print we promised. The chef repeats the same line, and the host spots twins. That's live and unedited, on Google's own channel.

So here's the template, in five parts:

• World: where you are and what's in it.
• Camera: exactly where the viewer stands and how they move.
• Look: the film stock or visual style.
• Rules: what must never happen, like the CGI look.
• Sound: what you should hear.

The catch: can you use it, and how fast is it?

The question with the most likes under Google's launch video was "when can I use this?" Right now, it's trusted testers only. Next, it's coming to Google AI Ultra subscribers inside Project Genie, and to developers through the Gemini API, a first for Genie. There's a waitlist. It doesn't run on your own computer, and it isn't open source.

Speed is about 24 frames a second. When the host asked about 60, the answer was "60 FPS, we never tried." So think film speed, not gaming speed. Google hasn't said how long a Genie 4 world lasts. In Genie 3's Project Genie, worlds lasted 60 seconds.

Every frame carries an invisible SynthID watermark, so Genie footage can be identified as AI-made. And when Genie 3 reached the public in January, Unity's stock fell by about a fifth in a single day. Game studios are watching this one closely.

Who's already using it

• Filmmakers run rough takes before they hire a crew.
• Journalists turn a single archival photo into a place you can walk through.
• Game designers test an idea before writing a single line of engine code.
• Robots practise in worlds where mistakes cost nothing.

Back to the crane. Six words, and the sky changed. The fine print is real. But so is the magic.

Sources: Google DeepMind's Genie 4 post, Google DeepMind's launch video, and the Google for Developers Release Notes interview "Introducing Genie 4" with Raia Hadsell, Diego Rivas and Hang Qi. The Unity share move is from Reuters and PocketGamer.biz, January 30, 2026.

Get the Genie Prompt Kit

We put all 12 official Genie 4 prompts in one PDF, word for word, with every paragraph tagged World, Camera, Look, Rules or Sound, plus the 5-part template and the fine print. Comment GENIE on the video, or grab it below.

How we counted: prompt word counts are ours, from the text Google published, not counting the "Prompt:" labels. All Genie footage in the video is Google's, shown with credit for commentary. Independent breakdown, not affiliated with Google.