Google DeepMind's Genie 4 AI builds virtual worlds you can edit with a sentence

A scene plays out on the screen. You type “make it rain” or “add a character,” and the world changes without stopping. On October 27, Google DeepMind unveiled Genie 4, the fourth generation of its “world model”: an AI that does not produce a fixed video but a simulated environment, where you can move around and interact in real time.

It is still only a research preview, limited to selected testers. According to TechCrunch, access will open in the coming weeks to subscribers of the Google AI Ultra plan, through the Project Genie service, and to partners of the Gemini API and AI Studio. A waitlist is open. The announcement does not mention any region; Project Genie has been rolling out gradually since May 2026 to eligible Ultra subscribers aged 18 and over worldwide.

What changes compared with Genie 3

The most visible novelty is sound. DeepMind presents Genie 4 as its first real-time world model to natively generate music, ambient noise and dialogue that adapt to events in the scene. According to TechCrunch, a scene stays coherent for up to two minutes, twice as long as with Genie 3, released a little over a year earlier. Google also highlights more believable physics (collisions, aerodynamics) and more natural human movements.

Genie 4 does not work alone. Gemini models interpret the user's requests, anchor the scene in a realistic look and break a goal down into steps. DeepMind names Gemini 3.8 Flash in this guide role, which checks progress and revises its plan when the world changes. In a robotics experiment, the Gemini 4 Argon model breaks down a task that Genie 4 then simulates, a way to test a robot before releasing it into the real world.

What could it be used for?

Google suggests several uses: film previsualization, where a director chains together “takes” in a simulation; rapid prototyping of interactive experiences; recreating a historical environment from an archive photo; simulating floods or fires; and training robots and AI agents in safe environments. Diego Rivas, an engineer at Google DeepMind quoted by TechCrunch, sees the tool as useful in the brainstorming and storyboarding phase, for coming up with ideas quickly. He acknowledges that it is still early and that the full product experience is not yet known.

Limits that Google acknowledges

DeepMind itself describes Genie 4 as an “early and experimental” model. Visual details and object permanence drift over time, especially in long sessions or after several edits. Complex physical interactions, sound accuracy, the legibility of text in the scene and long-term coherence remain open challenges. All images and sounds produced carry Google's invisible SynthID watermark.

Without independent evaluation, these claims currently rest on demonstrations chosen by Google. The real test will come with the opening to subscribers, when we see how the model holds up outside the launch examples.