Google DeepMind released Genie 4 on Monday as a research preview for trusted testers, and the headline change is not resolution or frame rate. It is sound.
The model is the studio’s first real-time world simulator to generate audio natively, according to the DeepMind blog. Every world it builds carries its own ambient bed, from wind moving through trees to surf breaking on a shore. Characters speak too, with dialogue that shifts as the user changes the scene.
Raia Hadsell, Diego Rivas and Hang Qi authored the announcement. Their framing puts world simulation near the center of the road to general intelligence, on the theory that an agent which can predict how an environment responds is better prepared for the one it actually inhabits.
Driving the system is the Gemini family. Gemini Omni grounds each scene in plausible real-world appearance, while other Gemini models handle the reasoning that keeps a simulation coherent across edits. A user names an event, and Genie 4 works out what happens next, including how the people in frame move and react.
DeepMind points to people as the biggest leap. Figures now walk with weight, respond to what surrounds them, and can be redirected mid-scene. Gemini 3.8 Flash sits in the loop as a planning agent, breaking multi-step goals into stages and checking incoming frames to confirm each one landed.
The studio also tested Genie 4 against Gemini Robotics, simulating an Apptronik Apollo humanoid working through a game of checkers.
Limits remain. Object permanence drifts over longer sessions or after repeated prompt edits, and legible text rendering is still unsolved. Every clip carries a SynthID watermark.
Access widens soon to Google AI Ultra subscribers through Project Genie, with a waitlist already open.
