For three weeks, Genie 4 existed only as a research preview that most people had to take Google's word for. That changed this week, when Google DeepMind folded the model into Project Genie and opened it to AI Ultra subscribers, turning a demo reel into something a paying customer can actually sit inside. The pitch is disarmingly simple: describe a place, step into a generated version of it, and then keep talking to reshape it while you're still standing there.
What makes that different from the wave of text-to-video tools that preceded it is persistence. Earlier world models could generate a few seconds of a coherent scene before the illusion collapsed into visual noise. Genie 4 holds a world together for up to two minutes, and users can intervene mid-session — summon a character, swap the weather, swing the camera around a building — without the simulation forgetting what it built a moment earlier. Google has paired that with real-time native audio that tracks the action, so a storm that rolls in on command also sounds like one.
The ambition behind this is bigger than entertainment demos. Google has been explicit that world models are a step toward systems that reason about physical space and cause-and-effect the way a robot or an autonomous agent would need to, rather than just predicting the next pixel in a video. A model that can simulate "what happens if I open this door" is a model that has learned something closer to physics than to cinematography, and that is the capability Google is racing Genie toward with each version.
None of that means the illusion is seamless yet. Google itself has flagged camera latency, fumbled physical interactions, and scenes that visibly drift after a long run or a complicated edit — the kind of seams that remind you a world model is still inferring its surroundings frame by frame rather than simulating them outright. Longer sessions make the cracks more visible, which is precisely why Google capped this version at two minutes before loosening the leash further.
What happens next matters more than the launch itself. Opening Genie 4 to subscribers rather than a closed set of testers hands Google something it cannot generate internally: millions of unpredictable prompts from people trying to break the thing in ways a lab never would. Every "world" a subscriber builds and discards becomes a data point about where persistence fails and where it holds — exactly the kind of pressure-testing that turns an impressive preview into a product that survives contact with the public.