Google DeepMind's Genie 4 adds sound to real-time worlds

Genie 4 builds a world you can steer by typing, and now you can hear it. You can't use it yet.

Sound, steering and a planning agent

Google DeepMind announced Genie 4 on Oct. 27, according to its blog post. It is a world model that generates a scene in real time and lets you steer it in natural language: change the weather, move the camera, add people or objects.

DeepMind calls it "our first real-time world model to generate audio natively." The post says it produces ambient sound to match each setting, and characters who hold real-time dialogue.

DeepMind says Gemini Omni anchors how the world looks and behaves, and that other Gemini models handle reasoning so scenes stay coherent.

The post also describes a planning agent built on Gemini 3.8 Flash. It watches the scene, breaks a goal into steps, checks frames to confirm progress and revises its plan when the world changes or you step in.

DeepMind says the team also tried Genie 4 for robotics research, including a checkers simulation with Apptronik's Apollo humanoid.

A research preview with no numbers

Access is narrow. "We have started rolling Genie 4 out as a research preview for trusted testers," DeepMind says.

Next, it says, Genie 4 is "coming soon" to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and Google AI Studio. No date.

The post gives no resolution, frame rate, maximum session length or benchmark. It does list its own limits. Visual detail and object permanence can drift, especially in longer sessions, after several prompt edits or while moving through a scene. Legible text, spatial accuracy and long-term consistency are still research areas, DeepMind says.

All video and audio Genie 4 generates will carry SynthID, Google's AI watermark, per the post.

The agent loop is the part to watch

Yes, it's another demo you can't touch. The audio is the headline. The more useful idea for builders is the loop: a model that plans inside a simulated world, checks its own progress and adjusts.

That is how you would train a robot or an agent cheaply before it touches anything real. A simulator that responds to plain language makes those test worlds faster to build.

DeepMind says its testers include creatives using Genie 4 for previsualization, journalists turning archival photos into simulations, developers prototyping games and enterprises testing scenario generation. If your work looks like one of those, join the Project Genie waitlist now so you're early when access opens.

If your work doesn't, there is nothing to do this week.

Watch the Ultra rollout and session length

Watch for the Google AI Ultra rollout date and for Gemini API pricing. Then watch how long a Genie 4 session holds together before the world starts to drift.

Until then, it's a waitlist.