Google Unveils Genie 4: AI Worlds You Can Explore, Hear and Steer in Real Time

Google DeepMind just announced Genie 4, a new AI model that turns a text prompt or a photo into a world you can walk through, hear and change while you are inside it. As the Google DeepMind blog explains, it is a big step up from Genie 3, with much more realistic scenes, sound generated on the fly and the ability to rewrite what is happening just by typing. It is in testing for now, and Google AI Ultra subscribers are next in line.

1. What Genie 4 actually does

Worlds, not videos: a video model like Veo makes a clip you watch. Genie makes a world you control. You describe a place (or give it an image), and Genie 4 builds it around you as you move, frame by frame, using standard keyboard keys for walking and arrow keys for the camera.

Set something in motion: Google says you can start an event and Genie 4 works out what happens next: how the world reacts, how it sounds, and how the people in it move and speak.

2. The four big upgrades

More realistic worlds: scenes look far closer to real footage than Genie 3, and people are the biggest jump. Characters move more naturally and react to what is around them.

Sound is generated too: Genie 4 is DeepMind’s first real-time world model that makes its own audio. That includes background sounds, music and dialogue that changes based on what you do.

Change the scene live: type a prompt while you are still exploring, for example asking for rain or a new character, and Gemini adjusts your request so it fits the scene. Between your prompts, Gemini also keeps the world moving in a sensible direction.

Longer sessions: according to TechCrunch, a scene can now stay consistent for up to about two minutes, double what Genie 3 managed.

3. Gemini is doing a lot of the work

Two Gemini models help: Google says Gemini Omni anchors each world in how things really look and behave, while Gemini 3.8 Flash acts as a guide in the loop. It watches each new frame, breaks goals into steps and checks that each step actually happened.

Robots in a pretend lab: DeepMind also showed Gemini 4 Argon planning a task step by step while Genie 4 renders it, including a humanoid robot playing checkers. The idea is to train robots and AI agents safely in made-up worlds before they touch the real one.

4. Who can try Genie 4, and when

Want in early? DeepMind has opened a waitlist for Genie 4 in Project Genie, linked from its announcement post. Project Genie is Google’s web app for building and exploring these worlds, and today it runs on Genie 3 for Ultra subscribers aged 18 and over.

On a Chromebook or Googlebook: Project Genie runs in the browser and the worlds stream from Google’s servers, so your laptop does not need a powerful chip. Once Genie 4 reaches your account, Chrome on a Chromebook or Googlebook should be all you need, though we will confirm that when we can test it.

5. What Genie 4 still gets wrong

Things drift: Google is upfront that details and objects can shift the longer a scene runs, after lots of prompt edits, or as you move around a lot.

Text and sound: readable text usually only appears if it was in your original description, and the accuracy of real-time audio is still being worked on.

Real places: Genie 4 cannot recreate real-world locations perfectly. Every video and sound clip it makes carries Google’s invisible SynthID watermark, so it can be identified as AI-made.

Watch Google’s Genie 4 video

In Google’s official video, Logan Kilpatrick talks with Google DeepMind’s Raia Hadsell, Diego Rivas and Hang Qi about Genie 4, with live demos from 5:10.

Our take

Genie 4 is one of the most impressive things Google has shown this year, and the jump from silent worlds to scenes with sound and talking characters is huge. For now it is a research toy for testers and, soon, for people paying for Google AI Ultra, so most of us will be watching demos rather than playing. Still, because it streams from the cloud, a cheap Chromebook should be just as good a window into these worlds as a high-end PC when it does open up.