DeepMind ships a world model that hears, bounces, and talks back
Genie 4 adds native audio, improved physics, and more realistic human behavior to DeepMind's interactive world simulation, rolling out as a research preview for AI Ultra subscribers.
On October 27, Google DeepMind announced Genie 4, the company’s newest interactive world model, a natively multimodal system that generates a simulated environment — including the audio that should inhabit it — in real time. The announcement, authored by Raia Hadsell, Diego Rivas, and Hang Qi, frames Genie 4 as a new generation of world simulation built on Gemini Omni and other Gemini models to maintain coherence across the visual, physical, and sonic layers of a scene as the user steers it live.
The audio is the headline shift. Where a world model previously implied a silent, purely visual simulation — geometry, lighting, textures, physics — Genie 4 generates ambient sound, music, and interactive dialogue natively alongside the world rather than bolting a soundtrack on afterward. A ball rolling across a wooden floor produces a different sound than the same ball on gravel; a character turning to address the user triggers speech synchronized to the generated audio stream. The system treats sound as a first-class output, not a post-production layer.
Physics and human behavior are the other two axes where Genie 4 distinguishes itself from its predecessor: aerodynamics, collision response, and the gait, posture, and vocal delivery of simulated people are all notably improved over Genie 3. The practical upshot is that a scene in which a generated character picks up a glass, walks to a window, and speaks to the user should now hold together across the visual, physical, and acoustic domains without the seams that were visible a generation ago.
Access is deliberately narrow. The system is rolling out as a research preview for trusted testers, starting with Google AI Ultra subscribers at $100 per month through Project Genie, alongside early-access partners building on the Gemini API and AI Studio. There is no fixed consumer rollout timeline — DeepMind is explicit that this is a preview for a small group of people stress-testing the model’s coherence under interactive conditions, not a feature landing in a phone this quarter.
What’s interesting about the trajectory from Genie 3 to Genie 4 is the compression of what a single forward-pass has to produce. A world model that must simultaneously maintain 3D geometric consistency, physically plausible dynamics, coherent character behavior, and a temporally aligned audio track — all while responding to real-time user input — is solving a substantially different problem than a video generator that renders pretty frames in sequence. The fact that DeepMind is gating this behind a $100 tier rather than shipping it broadly suggests they still consider coherence under open-ended interaction to be the unsolved part.