Google DeepMind announced Genie 4 on October 27, a world model for real-time interactive simulation. The post is signed by Raia Hadsell, Diego Rivas and Hang Qi. It is going out as a research preview to trusted testers first.
What Genie 4 does differently
Genie 4 is natively multimodal. It generates the world and its audio in the same pass: ambient sound, music, and dialogue that responds to what you do. In models like this, sound has usually arrived afterwards, or not at all.
The second difference is the steering. While a scene is running, you can change it with language: switch the weather, move the camera, add a person or an object. Gemini adapts your prompt to the surrounding scene instead of executing it literally. And between two inputs, Gemini watches and steers the scene itself so the world stays coherent.
Underneath sits the Gemini family, split into two roles. Gemini Omni anchors appearance and behavior in the real world, while other Gemini models supply the reasoning that keeps a simulation consistent over time. The Verge names Gemini 3.8 Flash as the base.
Who gets it
Right now: trusted testers. After that, Google AI Ultra subscribers in Project Genie, plus early access partners through the Gemini API and Google AI Studio. Project Genie has a waitlist. There is no price anywhere, and no figure for resolution or how long a scene can run.
The point is training grounds, not demos
DeepMind frames Genie 4 as a step toward AGI, and the argument is not about graphics: endlessly variable, safe environments where robots and agents can train. That is where this gets interesting for people working with Claude. Anthropic has no world model and is not chasing one; the overlap sits elsewhere, in the question of what an agent is supposed to practice on. Anyone testing agents today assembles those environments by hand. Genie 4 is the first serious attempt to generate them.