Google DeepMind Announces Genie 4, a Real-Time World Model With Native Audio
Google DeepMind announced Genie 4 on 27 October 2026, describing it as a new generation of real-time interactive world simulation. You type or steer your way through a generated world, and for the first time in the Genie line the model generates audio natively, including music, ambient sound and dialogue, alongside the picture.
DeepMind says Genie 4 simulates physics such as aerodynamics and collisions, produces more natural human motion, and can be steered in natural language while you move with standard keyboard controls. Gemini models act as the planner, helping keep scenes coherent and reasoning about what should happen next. Prompts can be changed mid-session, so you can alter the weather, the camera or the objects in a scene as you go.
Access is limited. Genie 4 is a research preview for trusted testers, with Google AI Ultra subscribers coming next through Project Genie, plus early access partners in the Gemini API and Google AI Studio. A waitlist is open. DeepMind has not published resolution, frame rate or session length figures in its announcement.
It is also candid about the limits. Visual detail and object permanence can drift over longer sessions or after several prompt edits, and complex physical interactions, legible text and spatial accuracy are still active problem areas.
For people who work in 3D and real-time, the relevant use is the one DeepMind lists for its testers: previsualisation and prototyping interactive experiences. A world you can walk through in seconds is a fast way to test mood, pacing and composition before anyone builds a single asset. It will not replace a modelled scene that has to hold up to scrutiny, because the consistency is not there yet.
That is how we would use it, if and when we get access. It is a sketchbook for atmosphere and layout, in the same spirit as a paper sketch or an animatic, and the real build still happens in proper tools. We will be watching how fast the drift problem improves.