Google DeepMind's Genie 4 lets you walk into a prompt and hear it back
The new world model generates photoreal scenes with sound in real time, steered by typed prompts and guided by Gemini. It is a research preview, not a product.
Key points
- Google DeepMind announced Genie 4 alongside enterprise testing of Gemini 4 Argon, pairing a reasoning model with a simulator from the same family.
- Genie 4 is a world model: software that generates a moving, interactive scene from a text prompt, now with ambient sound and spoken dialogue.
- Access goes first to trusted testers, then to Google AI Ultra subscribers through Project Genie and early-access developers on the Gemini API.
- Every clip Genie 4 produces carries SynthID, Google's invisible watermark for AI-made images, video and audio.
- Gemini 3.8 Flash runs in the background as a planner, breaking a user's goal into steps and checking each frame to confirm the simulation followed them.
Google DeepMind has shown off a system that turns a typed sentence into a small, playable world. It's called Genie 4, and the lab is pitching it as the next step toward machines that understand how physical reality actually behaves.
You type a prompt. A scene appears. Navigate it with the keyboard, type again to change the weather or add a person, and the scene adjusts while you watch. It now has sound too: wind, waves, footsteps, voices that reply when you address them.
That last part is new. Genie 3 was silent.
What is a world model, in plain words?
A world model is software that tries to simulate how a place looks and sounds and reacts, rather than just predict the next word or paint a still picture. Think of it as a game engine that writes itself from a sentence.
Genie 4 leans on two other Google systems to do this. Gemini Omni, the company's main multimodal model (one that handles text, images and audio together), grounds the look and behaviour of objects. Separate Gemini models handle the reasoning that keeps the scene consistent when you leave a room and come back.
The lab says the sharpest improvement since the last version is in people. Characters move more naturally, respond to their surroundings, and hold short conversations that adapt to what you type.
Who gets to use it, and when?
Almost nobody, yet. Google DeepMind is calling this a research preview and handing it first to trusted testers in film, journalism, education and engineering. After that, access opens to subscribers of Google AI Ultra, the company's top paid tier, inside a sandbox called Project Genie, and to selected developers on the Gemini API and Google AI Studio.
There's a waitlist. No consumer launch date has been given.
How does Gemini fit in?
Genie 4 does the pixels. Gemini does the thinking. A lightweight model called Gemini 3.8 Flash watches the simulation as it runs, breaks a goal into steps, and checks each frame to see whether the step worked.
DeepMind has also been pairing Genie 4 with Gemini 4 Argon, the enterprise-focused reasoning model AI2Day covered when it launched for regulated industries last month. In one demo, Argon planned the moves for Apptronik's Apollo humanoid robot playing checkers, and Genie 4 rendered the game around it.
This is the pattern worth watching. Google is quietly stitching its reasoning model, its simulator and its multimodal layer into one loop. If it holds together outside demos, that loop is how you train a robot or an agent without crashing a real one.
What are the testers actually doing with it?
Four use cases came up in the preview notes.
| Field | Use |
|---|---|
| Film and TV | Running multiple "takes" of a scene, changing light and dialogue live |
| Photojournalism | Turning a single archival photograph into a walkable environment |
| Game design | Prototyping a level before any code is written |
| Enterprise simulation | Generating training scenarios for robots and software agents |
Should ordinary people worry about fakes?
A bit, but Google has at least labelled the output. Every video and audio clip Genie 4 generates carries SynthID, an invisible watermark Google uses across its generative models. We've followed SynthID since it first appeared on 16 July 2026, and it keeps turning up in new places: the live avatar product in September now stamps every frame with it too. That doesn't stop misuse, but it gives platforms and fact-checkers a signal to look for.
For now, nobody can type a sentence and generate a convincing fake newsreel on their laptop. Access is gated. That'll change.
The honest verdict: watermarking is the floor, not the ceiling. Google's pattern this year has been to ship the capability first and explain the safety work after. Genie 4 fits that pattern exactly, and the audio dimension makes it a more pressing concern than the image watermarking debates that came before.
