Google DeepMind on Tuesday unveiled Genie 4, a new "world model" that generates fully interactive 3D environments in real time from a text prompt, adding native audio and more realistic people to a technology the company has cast as a building block toward artificial general intelligence.
Unlike a video game, which relies on pre-built environments and programmed logic, a world model generates its surroundings frame by frame as a user moves through them, reacting to natural-language instructions on the fly. According to DeepMind's announcement, users can change the weather, move the camera, or add people and objects simply by describing what they want, and the system adapts the scene to match.
Sound and physics, generated together
DeepMind says Genie 4 is the first real-time world model to generate audio natively alongside its visuals, producing music, ambient noise and dialogue that adapts to what is happening on screen. The model draws on Google's Gemini family throughout: Gemini Omni anchors how people and objects look and behave, while Gemini 3.8 Flash acts as a guiding agent that plans multi-step tasks, checks progress and adjusts the scene as it unfolds. The system can simulate physical interactions such as collisions and aerodynamics, and DeepMind says characters now move and react to their surroundings more naturally than in Genie 3, the version it unveiled last year.
DeepMind frames the technology as a tool for creating "infinite training environments" for physical robots, or for rapid prototyping in creative industries, rather than as a game engine — though the company's Genie line has repeatedly unsettled the gaming industry since Genie 3's debut sent some gaming stocks lower last year. The company acknowledges real limits: worlds stay coherent over short sessions, but visual detail and object permanence can drift the longer a session runs.
A staged rollout
Access to Genie 4 will begin as a research preview for "trusted testers," rolling out first to Google AI Ultra subscribers through the company's Project Genie prototype and to early-access partners of the Gemini API and Google AI Studio. A broader release to consumers and developers is planned for later, with a waitlist open through Project Genie in the meantime.
The announcement arrived in the same week Anthropic filed paperwork for what could be the largest stock listing in history, underscoring how central generative AI products have become to the valuations of the handful of companies racing to build them — even as researchers at DeepMind and elsewhere continue to debate how much of what world models produce should be read as progress toward general intelligence, rather than an increasingly convincing illusion of it.
