DeepMind's Genie 4 adds sound and live steering to AI worlds

Google DeepMind has announced Genie 4, the next version of its "world model" — an AI system that generates an interactive environment you can move through in real time, rather than a fixed video clip. The headline additions are native sound and live steering with text prompts, according to DeepMind's announcement on 27 October.

What's new in Genie 4

DeepMind describes four main advances:

• Real-time steering. While a world is running, you can type instructions — change the weather, move the camera, introduce a new person or object — and Gemini adapts the prompt to fit the scene. Between your inputs, Gemini also monitors the world to keep it progressing consistently. Navigation uses ordinary keyboard controls.
• More realistic physics and people. DeepMind says Genie 4 simulates effects such as aerodynamics and collisions, and that its biggest leap is in human characters that move naturally and react to their surroundings.
• Native audio. This is the company's first real-time world model to generate sound alongside the picture: wind in trees, waves, music and interactive dialogue, where characters talk and respond to you.
• Multi-step events. Gemini 3.8 Flash acts as a guiding agent: it breaks a goal into steps, watches the generated frames and checks whether each step worked, adjusting the plan if the user intervenes.

DeepMind also showed a research demo in which Gemini 4 Argon plans a task — Apptronik's Apollo humanoid playing checkers — and feeds the steps to Genie 4 as captions, an early look at using generated worlds to rehearse robot behaviour.

Who can use it

For now, Genie 4 is a research preview for trusted testers. DeepMind 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; there is a waitlist. No pricing or wider release date has been given. All generated video and audio carry SynthID watermarks.

The limits DeepMind admits

The company calls Genie 4 an early, experimental model. Scenes stay coherent over short sessions, but object permanence and visual details can drift over time, especially in longer runs or after many prompt edits. Complex physical interactions, audio accuracy, legible text, spatial accuracy and long-term consistency are still being worked on.

Why it matters

Testers are using it for film pre-visualisation, turning archive photos into explorable scenes, and prototyping games before writing engine code. For AI labs, world models are also training grounds for robots and agents. To understand the hardware side of running AI yourself, see our guides to how much VRAM you need for local AI and what an NPU is; for models you can download rather than rent, read what open-weight AI models are.