Google DeepMind’s Genie 4 is designed to make AI-generated worlds feel more coherent and lifelike, with improved motion, more natural-looking people and sound built into the experience. Users can describe a world and explore it as the model generates it frame by frame, then change details such as the weather or camera as they go.
Worlds that respond as you explore
A world model creates an interactive environment without requiring developers to pre-render each setting or program every interaction. A prompt gives the system a starting point, and the environment takes shape in real time as the user moves through it.
That approach can make generated worlds useful for more than entertainment. DeepMind says they could provide training environments for physical robots or help creative teams rapidly prototype ideas. Those uses depend on whether a simulation behaves convincingly enough to be useful.
Genie 4 uses Gemini 3.8 Flash as a guiding agent. It breaks down goals and evaluates the frames the model produces, helping preserve the prompt’s context as the environment changes. DeepMind says the worlds can still lose consistency after a couple of minutes, so long-term continuity remains a limitation.
More believable motion and people
DeepMind says Genie 4 makes progress on physics that can look wrong in generated video. Its simulations are intended to handle aerodynamics and collisions more realistically, helping objects and events feel more plausible as the scene unfolds.
People in the generated environments are also meant to feel less uncanny. The company says they can move more naturally and respond to changes in their surroundings. If those improvements hold up in use, they could make simulations more convincing and more useful for exploring scenarios.
Users can also adjust a simulation while it is running. They might change the weather, introduce new elements or move the camera, with Gemini helping guide the result after each update. That makes the environment a continuing exchange: the user can steer it instead of accepting a single generated scene.
Audio makes the worlds more immersive
Genie 4 is DeepMind’s first world model with native audio. Generated videos can include ambient background sounds, and prompts can specify audio elements. The model is also intended to produce interactive dialogue that changes as the simulation changes.
Speech that responds to events could make an environment feel more like an active place than a sequence of images. DeepMind is presenting this as part of the model’s capabilities, though the source does not describe how well dialogue works in practice.
Access will expand in stages
World models remain an active research area. Genie 3 did not expand beyond the limited Project Genie web experience, and DeepMind says Genie 4 has so far been tested with a small group of trusted users. Wider access is expected to begin expanding soon, but there is no specific timeline.
Developers are expected to get access through the Gemini API and AI Studio, with pricing not yet available. Consumer access will begin in phases through Google AI Ultra plans, which start at $100 per month. Google may eventually extend access to AI Pro and AI Plus plans.
Consumers will use Genie 4 through Project Genie, which will limit how much they can do with the model. People interested in early access can join a waitlist. The planned rollout gives developers and some subscribers a route to try the system, while its short consistency window and staged availability show that the experience is still developing.