Google DeepMind's Genie 4 builds explorable worlds with sound
On 27 October 2026, Google DeepMind announced Genie 4, an AI that makes a whole world you can walk around in and change as you go. Earlier tools of this kind produced a picture or a video.
This post explains what that means, what Genie 4 can do, and why most people cannot try it yet.
What is Genie 4?
DeepMind calls Genie 4 a world model. In plain words, a world model is an AI that builds a place on screen and works out what happens in it from moment to moment.
A video game is a useful comparison. In a normal game, people design every tree, wall and sound in advance. With Genie 4, nobody builds the world ahead of time. The AI creates it on the spot, while you are exploring it.
It also works in real time, so the world responds straight away when you do something. You are not waiting for a finished video to be produced.
What can it do?
You move around a Genie 4 world with ordinary keyboard controls, the same kind you might use in a computer game. The keys let you walk through the scene and point the camera where you want to look.
You can also type instructions in everyday language to change the world while you are in it. DeepMind gives examples such as:
- changing the weather
- moving the camera
- adding a new person or object to the scene
When you type a request, the AI adjusts it to fit what is already there, so a new person should look like they belong in that place. Between your instructions, it keeps track of the scene so the world goes on making sense.
The biggest new feature is sound. DeepMind says Genie 4 is its first real-time world model to create audio itself. Each world comes with the background noise of its setting, such as wind in the trees or waves at the beach, and it can make music too.
Characters in a Genie 4 world can also talk, and what they say changes depending on what you do or what happens around them. DeepMind describes human speech as "the real frontier" here.
DeepMind also says the pictures look more lifelike than before and that people in the worlds move more naturally. According to the company, the model copies real-world physics, the rules for how things behave, such as how objects crash into each other and how things move through the air.
The demonstrations DeepMind shared cover a wide range. They include a man playing a saxophone on a platform floating on a misty lake, a falling maple seed, a busy Japanese train station from the 1980s, a hot air balloon ride over Antarctica, a landscape made entirely of stained glass, and a race on the back of a dragon. DeepMind says Genie 4 produced all the sounds in its examples itself.
How does it work?
Genie 4 builds on Gemini, Google's family of AI models, with different models doing different jobs.
One, called Gemini Omni, keeps the simulation grounded in how things really look and behave. Other Gemini models handle the reasoning needed to keep the world consistent from one moment to the next, and they also adapt your typed requests to suit the scene.
According to DeepMind, this means you can set something in motion and Genie 4 works out what happens next: how the world responds, how it sounds, and how the people in it move, react and speak.
Why does it matter?
DeepMind says understanding and simulating the world is an important part of its long-term AI research.
It points to practical uses as well. These worlds can be made safely and in endless variety, so they could serve as training grounds for robots and for AI agents, which are AI programs that carry out tasks on their own. A robot could practise in a simulated world before trying something in the real one.
DeepMind also says Genie 4 could lead to new kinds of interactive experiences. If you run a small business or simply have an interest in games, storytelling or entertainment, that's the part I'd keep an eye on.
Can you try it yet?
Only if you are invited. Genie 4 is currently a research preview, an early version shared with a small group of trusted testers, and DeepMind is collecting their feedback first.
You can join a waitlist through Project Genie, Google's space for trying Genie.
After the testing stage, DeepMind plans to open it up to Google AI Ultra subscribers in Project Genie. It also plans to offer it to early access partners through the Gemini API and Google AI Studio, tools that let developers build Google's AI into their own products.
A lot is still unknown. DeepMind has not shared any test scores comparing Genie 4 with other systems, so its claims about realism cannot yet be checked independently, and it has not said when the wider release will happen. For now, Genie 4 is an impressive preview that most people cannot use.
Source: Genie 4, Google DeepMind. Listed in AI News Week 45, where we track every AI release each week.