Genie 4 is DeepMind's world model you can steer mid-scene, with sound

Google DeepMind announced Genie 4, the next version of its "world model": you describe a place (or hand it a reference image), and it generates an explorable scene frame by frame, in real time, that you move through with keyboard controls. Nothing is pre-built. There is no game engine and no level file; the model makes up what you see as you go.

What's new, according to DeepMind.

• Change the world while it runs. You type things like a weather change, a camera move or a new person or object, and the scene adapts. DeepMind says Gemini rewrites your prompt to fit the scene and steers the world between your inputs so it stays coherent.
• Sound, including speech. DeepMind calls Genie 4 its first real-time world model with native audio: ambient sound, music and characters who talk and react to you. It says all the audio in its demo videos came straight from the model.
• Gemini in the loop. Gemini Omni anchors how things look and behave, and Gemini 3.8 Flash acts as a "guiding agent": it watches the scene, breaks a goal into steps and checks frames to confirm each step happened. DeepMind also shows Gemini 4 Argon planning a robot task (Apptronik's Apollo humanoid playing checkers) that Genie 4 renders.
• More realistic physics and people. DeepMind says it simulates things like aerodynamics and collisions better, and that people move and respond more naturally. It published demo clips, not benchmark numbers.

Who it's for. DeepMind's trusted testers are using it as a "draft mode" for film pre-visualization, for turning an archival photo into an explorable scene, for game prototyping "before writing a single line of engine code," and for generating simulation scenarios, including robot training.

Who can use it. Not most of us, yet. Genie 4 is a research preview for trusted testers. Google says it is "coming soon" to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio, and there is a waitlist. Google has not announced a price or a date for the API.

The catch. Google lists the limits itself:

• Visual details and object permanence "can shift or drift over time," more so in longer sessions, after several complex edits, or as you move around.
• The model supports "a few minutes of continuous interaction, rather than extended sessions."
• It can't simulate real-world locations accurately, legible text usually only shows up if it was in your prompt, and real-time audio accuracy is still being worked on.
• Everything it generates carries Google's SynthID watermark, which Google says helps identify it as AI-made.

Our take. Changing a scene with text while it runs, with sound, is the part that makes this feel like a tool instead of a demo reel. For game designers and previz teams, "idea to walkable prototype in seconds" is a real pitch. But a few minutes per session, with drift, means it's for sketching ideas, not shipping them. Wait for API access and pricing before you plan anything around it.

What we did and didn't do: we read DeepMind's announcement and Genie 4 model page and its launch post on X on October 27, 2026. We have not used Genie 4; the capability claims are Google's, and its demos are clips Google chose. Not sponsored; no affiliate links.