Google DeepMind announced Genie 4 on October 27 as a research preview for trusted testers, and the headline change is sound. The new world model generates explorable, interactive scenes in real time from a text prompt or an image, and for the first time it produces audio natively alongside the picture: ambient noise, music, and characters who can converse and react as events unfold. DeepMind says scenes now hold together for up to about two minutes, roughly double the limit reported for Genie 3, and every video and audio clip it emits carries a SynthID watermark.
The second change is control while the simulation is running. Users navigate with keyboard controls and can type natural-language prompts mid-scene to change the weather, move the camera, or introduce a new person or object, instead of starting over with a fresh prompt. Behind that sits a small stack of Google models: Gemini Omni anchors how objects look and behave, while Gemini 3.8 Flash serves as a guiding agent inside the loop, observing the scene, breaking goals into steps, and checking frames as Genie renders them. DeepMind also claims better physics, including aerodynamics and collisions, and calls its simulated people the most dramatic improvement.
For builders, the interesting part is the architecture, not the demo. A world model is becoming a system: a renderer that predicts the next frame, a reasoning model directing it, and provenance watermarking built into the output rather than bolted on afterward. That pattern points at practical uses DeepMind and early testers name, from film pre-visualization and game prototyping to training robots in simulated places before they touch the real world. Access, though, is gated. Trusted testers have it now; Google AI Ultra subscribers through Project Genie and early-access partners in the Gemini API and Google AI Studio come next, with no firm date, resolution, or pricing published.
DeepMind is also candid about the limits, and practitioners should take them at face value. Visual details and object permanence can still drift over time, especially in longer runs, after several complex prompt edits, or while navigating a scene. Legible in-world text, spatial accuracy, complex physics, and real-time audio accuracy remain open problems. Two coherent minutes is enough for a storyboard, a rehearsal, or a training episode. It is not a persistent world, and nobody should design a product that assumes it is one.
The same day underlined where the rest of the stack is moving. Oracle and Google announced that Google TPUs will run on Oracle Cloud Infrastructure from 2027, the first hyperscale cloud outside Google to offer them, with eighth-generation TPU 8t systems for training and TPU 8i systems for inference, while Oracle also integrates ChatGPT and Codex into its business applications. Models, chips, and distribution are all unbundling at once. Genie 4 fits that story: treat it as a serious preview, prototype against the pattern of a director model steering a simulator, and wait for published pricing and limits before committing a roadmap to it.
