Google DeepMind has unveiled Genie 4, a new world model that is being positioned as a sharper, more realistic successor to Genie 3. The update will roll out soon to developers and Google AI Ultra subscribers, with access set to expand through the Gemini API and AI Studio.
That timing matters because Google is not just showing another demo. It is pushing a model built to let users change the weather, add elements, move the camera and even set audio cues while the simulation runs, all while the system tries to keep the world coherent from one frame to the next.
DeepMind says Genie 4 improves on Genie 3 in almost every way, and the gains are not cosmetic. The model uses Gemini 3.8 Flash as a guiding agent to break down goals and judge each output frame, which gives it more control over what appears on screen and how the scene responds when prompts change midstream. It can better handle aerodynamics and collisions, and people in its simulations are meant to move and react in ways that look less uncanny.
The strongest change may be audio. Genie 4 is the first DeepMind world model with native sound, which means its videos can carry ambient background noise, user-specified audio elements and interactive dialogue that shifts as the simulation changes. For developers, that opens the door to richer test environments; for creative work, it could speed up rapid prototyping without waiting on fully built assets.
But the new realism still has a short leash. DeepMind says the worlds only stay consistent for a couple of minutes before they forget how things started. That limitation sits at the center of the pitch: the model is getting better at making believable spaces, yet it still cannot hold a world together for long enough to behave like a persistent one.
That gap helps explain why DeepMind keeps framing world models as a major research path, not a finished product. The company says they matter for the march toward artificial general intelligence and could one day create infinite training environments for robots or unlock new routes for creative prototyping. For now, Genie 4 is the latest step in that direction, not the end of it.
DeepMind has not given a specific timeline for wider access, and it has not spelled out what developers will pay. The exact pricing is still unclear, but the structure is easy enough to read: Google AI Ultra subscribers will get access alongside developers, and the broader rollout is set to begin soon through Gemini API and AI Studio as Google tests how far this more interactive world model can go.
