Google DeepMind launched Genie 4 on October 27, showing an AI model that generates worlds you can move through, reshape and hear in real time. The bigger ambition goes beyond games: Google wants simulation environments where AI agents and robots can learn before acting in the real world.
Unlike a fixed AI-generated video, Genie 4 produces a changing simulation. Testers navigate using keyboard controls and steer it with text prompts, from changing the weather to adding a person or object while a scene is running.
Worlds that react and speak
Genie 4 generates ambient sounds, music and dialogue alongside visuals. Google says characters can react to events and respond to a user's interactions as they happen. Its demonstrations also show simulated aerodynamics, collisions and human motion.
Google is using Gemini models to keep those worlds coherent. Gemini Omni helps anchor their appearance and behavior, while Gemini 3.8 Flash can guide an agent through a task, checking whether each step succeeded. The company has also tested Genie 4 with Gemini 4 Argon for robotics scenarios.
Watch Google DeepMind's official demonstration:
Who gets to try it
Genie 4 is initially available only to a limited group of trusted testers. Google says access is coming to Google AI Ultra subscribers through Project Genie, as well as early-access partners using the Gemini API and Google AI Studio. It has not given a date for a broad release.
Early testing spans filmmaking, journalism, education, gaming and engineering. Researchers and creators are exploring scene prototyping, historical reconstructions and scenario generation, though these are still experimental applications.
The limits still matter
Google acknowledges that objects and visual details can drift in longer sessions, especially after repeated changes. Complex physics, accurate real-time audio, readable text and long-term scene consistency remain difficult problems.
The company says Genie 4's generated video and audio will include SynthID watermarks, and that it conducted safety evaluations before opening the research preview.
Genie 4 points to a future where AI systems respond to actions rather than simply producing content. The harder test is whether those simulated worlds become reliable enough to help machines learn skills they can use outside a screen.
Get stories like this in your inbox every week. Subscribe to The Week in Perspective.
Sources: Google DeepMind's announcement and Google's launch post.