Google DeepMind's Genie 4 builds interactive worlds, now with sound

Google DeepMind presented Genie 4 on 27 October 2026. Genie is a so-called world model: an AI that creates a virtual world you can walk through while you watch. You control the camera and your character with the keyboard, and the AI keeps drawing the next image. No game engine and no pre-built world are needed.

What is new in Genie 4 is sound. According to DeepMind, the model is the first to create background sound, music and conversations at the same time as the images. So a busy street sounds like a busy street, and characters talk back when something happens.

What Genie 4 can do

According to DeepMind, Genie 4 imitates physics, such as air flows and collisions. Characters move naturally and react to their surroundings. You can also control them while you play.

You can change the world simply by typing what you want. Make it rain, move the camera or add a person or object. A Gemini model, Google's language AI, adapts such an instruction to what is already on screen. Along the way, it makes sure the world stays consistent.

For longer tasks with several steps, Genie 4 works together with Gemini 3.8 Flash. That model watches along, splits the goal into steps and checks in the new images whether each step has worked. If something goes differently than planned, or the user steps in, it adjusts its plan.

DeepMind is also testing a combination with Gemini 4 Argon and Gemini Robotics, Google's AI for robots. Argon splits a task into steps and passes them on to Genie 4 as short text instructions. In one of the examples, a simulated humanoid robot plays draughts. Videos of this and other examples, with sound, are on DeepMind's blog.

According to the tech site The Verge, a session can now last about two minutes. With the previous version, Genie 3, it was one minute.

Who can use it

Genie 4 is now available to try for a select group of testers. There is a waiting list. After that it will come to Google AI Ultra subscribers, in the Project Genie app, and to developers who have early access to the Gemini API and Google AI Studio.

All images and sound get an invisible watermark, SynthID. This makes it possible to check later that they were made by AI. DeepMind says it tested the model internally for safety beforehand.

According to DeepMind, testers are already using Genie 4 in different ways. Filmmakers use it to try out scenes quickly before they start filming. Photojournalists turn a single photo into a world you can walk through. Game makers test ideas for a game without writing code first.

Where it still needs to improve

DeepMind is open about what does not work well yet. In longer sessions, details can shift, and an object that was out of view for a moment sometimes looks different when it comes back. This happens mainly after many changes or when you walk around a lot. Complex physics, sound that matches exactly, readable text, correct distances and a world that stays the same for a long time are also still being researched. So far, all the examples come from Google itself. What the model can do will become clear as more people work with it.

What does this mean for the road to the singularity?

According to DeepMind, understanding and imitating the world is an important step towards AGI, AI that is as broadly intelligent as a human. An AI that can simulate a world with physics, sound and characters that react understands at least something of how that world works. Such a virtual world can also become a practice ground. Robots and AI assistants can try out endless situations there before they start work in the real world, as the robot example already shows. To do this, DeepMind wants to combine language models, which make plans, with Genie, which creates the images. If it succeeds in keeping the world stable in long sessions as well, that practice ground will come a lot closer.

What is the technological singularity? Read the explanation →