Genie 4 adds sound and live steering to Google’s world model, for testers only

Google DeepMind has introduced Genie 4, a new version of its world model that builds interactive, real-time environments with generated sound. Bloomberg reported the upgrade in the context of investor worries about video game companies. For now, Genie 4 is a research preview for trusted testers.

Sound, steering and a Gemini planner

According to DeepMind’s announcement, Genie 4 generates music, ambient sound and interactive dialogue together with the visuals. Users can change the weather or the camera angle, or add people and objects, by typing natural-language instructions while the simulation runs. The model adapts each instruction to the scene that is already on screen.

DeepMind also says the model simulates aerodynamics, collisions and characters that move and react in a natural way. Gemini Omni grounds the simulations in how real places look and behave, and other Gemini models help keep scenes coherent. For longer events, a Gemini model acts as a guiding agent: it plans steps, checks progress and adjusts when the user intervenes.

DeepMind describes the longer-term direction as hybrid world models, which pair language-based reasoning with pixel-level simulation. Testers have tried the model for creative pre-visualization, recreating historical scenes, game prototyping and general simulation.

Who can try it, and when

Access is limited. Genie 4 is a research preview for trusted testers. DeepMind says broader access is planned for Google AI Ultra subscribers through Project Genie, and for early-access partners in the Gemini API and Google AI Studio. A waitlist is open. The post does not give a date, pricing, regional availability or usage limits.

Google says the model ships with safety guardrails and SynthID watermarking. It also lists limitations. Visuals can drift and objects lose permanence over long sessions, and text rendering, spatial accuracy and physical interactions remain weak spots. Those are the company’s own descriptions, and no independent testing has been published in the sources we read.

A year of steady gains since Genie 3

Genie 4 arrives a little more than a year after Genie 3 appeared as a research preview, TechCrunch reported. Its report says Genie 4 holds a consistent scene for up to two minutes, double the limit of its predecessor. TechCrunch also quotes DeepMind engineer Diego Rivas saying it is “still early days for Genie 4”, and that real-time generation is most useful in ideation and storyboarding rather than final production.

That framing matters for the market angle. The Bloomberg headline ties the release to investor anxiety in game stocks, but two minutes of drifting, silent-until-now footage is not a game engine, and Rivas himself placed turning prototypes into finished assets, via Gemini Omni, as something that is “still a ways off”, according to TechCrunch. Nothing in the announcement claims Genie 4 replaces a game studio’s pipeline. The pitched uses are prototyping, previsualization, learning and robotics.

The Verge, as summarized by Techmeme, reports that Genie 4 is powered by Gemini 3.8 Flash. TechCrunch adds that a research experiment paired Genie 4 with Gemini Robotics, where a Gemini model broke tasks into steps that Genie 4 simulated in real time. Our own coverage of that Gemini family includes the earlier Gemini 4 Argon story, and the hardware side is covered in our piece on Google’s eighth generation TPUs.

What the announcement leaves open

Several things are not established. DeepMind gives no resolution or frame rate in the post, no benchmark against Genie 3, and no figures on how often drift or errors occur. The two-minute figure comes from TechCrunch, not from the blog text we could read. Until Ultra subscribers or API partners get hands-on access, the capabilities described here remain Google’s own claims.