Genie 4 Turns a Text Prompt Into a World You Can Walk Through, With Sound and Physics

Genie 4 Turns a Text Prompt Into a World You Can Walk Through, With Sound and Physics

6 min readOctober 29, 2026

Quick verdict

Google DeepMind shipped Genie 4, a world model that turns a text prompt into an environment you can move through in real time. The jump from Genie 3 is not subtle. This one generates its own ambient sound, music, and spoken dialogue, runs a physics pass for collisions and human motion, and keeps a Gemini model running in the loop to plan events and stop the scene falling apart. You cannot buy it yet. It is a research preview for trusted testers, with Google AI Ultra subscribers next in line and API access after that. But that last part is the real story: Genie is finally getting an API, which turns a lab demo into something developers can build on.

What actually shipped

Genie 4 is a step up on three fronts that Genie 3 either lacked or did badly.

  • Native audio. It generates ambient sound, music, and real-time dialogue with characters inside the scene. That is a first for Genie, and it is the difference between a moving postcard and something that feels like a place.
  • Physics and motion. DeepMind claims aerodynamics, collisions, and natural human movement. The team was blunt about the old version: Genie 3's worlds were too static. Genie 4 reacts to events you prompt while it is still generating, so you can drop something into the scene and watch the world respond.
  • A Gemini planner in the loop. The model is built on Gemini Omni, and Gemini runs during generation to plan multi-step events and hold the scene together over time. That is the part that keeps a long session from drifting into nonsense.

Early testers put it through its paces in public. Ethan Mollick described prompting a "neogothic brutalist city," then adding dragons and a talking NPC while the world was live. His verdict, echoed by others with access, was that it is not flawless but a large leap over Genie 3.

The robotics angle is the one to watch

DeepMind is not pitching Genie 4 only as a toy for building dreamlike cities. It frames the model as a way to make safe virtual environments for testing agents and robots, the kind of place you can let an agent fail a thousand times without breaking real hardware.

Two demos make the case. In one, a Gemini 3.8 Flash agent breaks a task into steps, checks each step, and replans as the generated scene changes around it. In another, a Gemini 4 Argon agent drives an Apollo humanoid playing checkers inside a Genie 4 world. Neither is a product. Both show the loop Google is after: a model that imagines an environment, and an agent that learns to act in it, with no physical robot at risk.

Who gets it, and when

Access rolls out in stages. The research preview is open to trusted testers now. Google AI Ultra subscribers get in next through Project Genie. After that, early-access partners reach it through the Gemini API and AI Studio, which is the first time Genie has been available through an API at all. There is a waitlist in the meantime.

That order tells you how experimental this still is. Google is gating the most capable version behind its top-tier subscription and a partner program rather than opening the floodgates, which is the same pattern we saw with Gemini 4 Argon's launch: announce the capability, then meter who can actually touch it.

Video: Genie 4 in motion

Static screenshots undersell a world model, so it helps to watch one generate and respond. Here is DeepMind's own walkthrough of what Genie 4 produces.

Why it matters if you pay for AI

For most people the honest answer today is that it does not change your week, because you cannot use it. The longer point is about where the frontier labs are spending their lead. A world model is a different thing from a video generator. A text-to-video tool hands you a finished clip; Genie 4 hands you a space you can steer, with physics and sound reacting as you go. That is the raw material for training and testing agents, and it is a place Google now has a visible head start.

It also widens the three-way race in a direction OpenAI and Anthropic have not matched in public. If you want to see how the labs line up across models rather than world simulators, our OpenAI vs Anthropic vs Google breakdown tracks that. And because capability keeps landing behind different subscriptions and waitlists, the practical move for anyone paying for several tools is to not lock into one vendor, which is the whole reason an AI aggregator exists.

FAQ

What is a world model, in plain terms?

It is a model that generates an interactive environment rather than a fixed image or clip. You give it a prompt, it builds a space, and you move through that space in real time while the model keeps generating what you would see and hear next.

How is Genie 4 different from a text-to-video generator?

A video generator produces a clip you watch. Genie 4 produces a world you act inside, with camera control, prompted events, physics, and audio that respond as you go. The output is interactive, not a finished render.

Can I use Genie 4 right now?

Not unless you are a trusted tester. It is a research preview, with Google AI Ultra subscribers next and API access for early-access partners after that. Everyone else joins a waitlist.

Why does Google keep mentioning robots?

Because a generated world is a cheap, safe place to train and test agents. You can let an agent or a simulated robot fail repeatedly inside Genie 4 without risking real hardware, which is hard to do any other way at scale.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.

Try Admix for $1.99

Related articles