Google unveils Genie 4: an AI model that generates interactive worlds with sound and living characters

What's new

Google DeepMind has released Genie 4, a new version of its world model technology. It's a neural network that doesn't draw a pre-made video but builds an interactive world on the fly: the user moves, pans the camera, types a text prompt—and the model reshapes the scene in response, frame by frame, in real time. Genie 4 is built on Gemini, and it's the first version in the lineup to add sound and characters who move, react to events, and speak, on top of the visual environment.

What changed since the previous version

The earlier Genie 3, unveiled in 2025, could only generate the visual side of the world—terrain, objects, camera movement. There was no sound or living characters. Genie 4 adds that layer: the simulation now includes how the world sounds and how people inside the scene behave—their movements, reactions, and speech. For a world model, this is a fundamental step: these systems used to be "silent" video generators with interactive controls; now they're a simulation with a full audio and social environment.

Controls in Genie 4 work on two levels:

• ordinary keys handle character navigation and camera control, just like in any game;
• text commands change the scene itself on the fly: you can add an object, trigger a new event, or change the weather or setting without leaving the current frame.

DeepMind specifically emphasizes that every demo clip in the announcement is a real-time interaction session with the model, not a pre-scripted, pre-rendered video. For generative demos from companies like this, that's an important detail—it's nearly impossible for a viewer to tell real interactivity from editing—and here Google is explicitly stating there's no editing: every frame shown is the model's response to a specific user action.

Who gets access, and when

Google is rolling out Genie 4 in stages rather than to everyone at once:

• the model first goes to trusted testers in a research preview—closed access for a limited group of users who provide feedback ahead of the wider launch;
• next in line are Google AI Ultra subscribers, Google's premium paid AI subscription; they'll get access to Genie 4 through the Project Genie interface, built specifically for entering generated worlds;
• in parallel, access will open to early Gemini API partners—developers who'll be able to embed Genie 4 in their own products;
• and to early-access participants in Google AI Studio, Google's developer environment for experimenting with its models.

The Project Genie waitlist is already open. Google hasn't given specific dates for the full launch or pricing—only promising that access will expand "soon."

Why it matters and who it's for

Interest in world models stems from practical utility, not just demo spectacle. For game studios, the technology makes it possible to rapidly prototype levels without manually coding the environment. For researchers training autonomous agents and robotics systems, it provides a nearly endless source of training environments, removing the need to manually build every simulation. And for creators of interactive content, it allows a scene to be assembled with words, without traditional 3D development.

Genie 4 being built on Gemini isn't incidental—it combines the model's language understanding with simulation of the physical world in a single system: the model doesn't just know what the world looks like, it also understands what logically should happen next given a text prompt.

That said, the announcement leaves open questions Google doesn't address: how long a world stays coherent before its logic starts to unravel, what the actual generation resolution is, and how much compute a single session requires. The answers to those questions will determine whether Genie 4 becomes a working tool for developers or remains an impressive technology demo.

What to do right now

If you work in games, training simulations, or agent development, it's worth joining the Project Genie waitlist and applying early for Gemini API access through Google AI Studio. While the model is still closed to the public, it's worth working out two or three concrete test scenarios in advance: how the model reacts to on-the-fly text commands, how the generated characters behave, and how convincing the audio environment sounds. That way, once access opens, you can quickly judge whether Genie 4 is ready for prototyping your particular use case. It's too early to build a finished product on it, but now is the time to start planning your testing.