Google DeepMind's Genie 4 world model: research preview, 2 minutes

Google DeepMind announced Genie 4 on 27 October 2026, describing it as a world model: software that simulates a scene in real time while a person steers it. It is a research preview, so it is available only to trusted testers for now. DeepMind says wider access starts with Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio. A waitlist is open.

What DeepMind says changed

• Native audio. DeepMind calls Genie 4 its first real-time world model to generate sound itself, from ambient noise to interactive dialogue in which characters respond to what the user does.
• Steering by text. Users can change the weather, move the camera or add a person or object mid-simulation, and Gemini adapts the prompt to the scene.
• Closer ties to Gemini. DeepMind says Gemini Omni anchors the look and behaviour of the world, and other Gemini models help keep the simulation coherent.
• Physics and people. DeepMind says the model simulates aerodynamics and collisions and moves human characters more naturally. These are the company's claims; the post does not cite an independent test.

What reporters add

TechCrunch reports that Genie 4 keeps a scene consistent for up to two minutes, double Genie 3's limit. It also reports that Gemini acts as a director that plans steps and checks frames while Genie renders them. In one research experiment tied to Gemini Robotics, a Gemini 4 Argon model split a task into steps that Genie 4 simulated live, which suggests a way to test robot behaviour before running it for real.

Google's research director Tejas Kulkarni told reporters the future is hybrid world models in which Gemini writes code and text and Genie simulates the result. DeepMind's Diego Rivas said the early users see value in storyboarding and rapid ideation, and stressed that it is early days and the end-to-end product is not settled, according to TechCrunch.

What a deployment team should take from it

Nothing here changes a rollout this quarter. There is no published price, no rate limit, no data-handling terms and no service commitment, and a two-minute consistency window limits it to short interactive sessions. The relevant signal is the direction: simulated environments are being pitched as a place to test agents and robots before they touch real systems, which is the same instinct as a staging environment for an agent.

If you evaluate agents today, keep your own test harness and treat generated worlds as one possible input later. If you buy through Google Cloud, ask your account team whether Genie 4 sits inside your existing data and retention terms before any employee uploads company images as reference frames.