DeepMind Gives a Few Testers a World Simulator That Makes Its Own Sound
Google DeepMind said on Oct. 27, 2026, that it has started releasing Genie 4, its new world model, to trusted testers as a research preview. A world model generates a scene a person can move around in, and DeepMind says this is the first of its real-time world models to produce a scene's sound at the same time as its picture.
DeepMind's announcement says the model is meant for building training environments for robots and software agents and for creating new interactive experiences, and that pairing it with Gemini lets the company test how AI systems handle an environment before they are used in the real world. Access stays narrow for now: Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio are listed as next in line, with a waitlist open in the meantime.
Two other changes are described. A user can type a plain instruction while a simulation runs, to change the weather or add a person or an object, and Gemini fits the request to the surrounding scene. A second model, Gemini 3.8 Flash, also sits in the loop: DeepMind says it watches the scene, breaks a goal into a sequence of steps, and reviews incoming frames to check that each step succeeded.
DeepMind's own list of what does not work yet is specific. Visual details and object permanence (whether things stay put) can shift or drift over time, especially as a simulation runs longer or after several complex prompt edits. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are all still active areas of research.
The demonstrations on the page are video DeepMind produced itself, and the announcement carries no benchmark score and no independent evaluation of Genie 4.
DeepMind says the testers are groups working in creative industries, journalism, education, gaming and engineering. Every video and audio clip Genie 4 produces will carry SynthID, the company's digital watermark for AI-generated content.
