DeepMind shows Genie 4 to trusted testers, with Gemini 3.8 Flash steering the world

Watch

Lab claim

On October 27, 2026, Google DeepMind announced Genie 4, a world model that generates a scene and its sound together. It is a research preview for trusted testers. DeepMind says Gemini 3.8 Flash watches the frames and checks that each step happened. The post does not give a frame rate or a resolution, and the model is not in the public API.

At a glance

Date	October 27, 2026
Access	Research preview for trusted testers
Later access	Google AI Ultra in Project Genie, then early-access API and AI Studio
Audio	Generated with the scene. DeepMind's first real-time world model to do this.
Look and behavior	Gemini Omni, on DeepMind's account
Checker	Gemini 3.8 Flash, in the loop
Frame rate and resolution	Not printed

How to read this claim

Measurement — The post does not print a frame rate or a resolution.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — The public API date is 'coming soon,' not a day.

Keep the habit · field card with prices and worked checks

What changed

Earlier Genie models simulated a world you could move through. Genie 4 adds native audio and a Gemini 3.8 Flash checker, and it is still a tester preview.

What it introduces

A world model that generates sound with the scene, still limited to trusted testers.

Who benefits Anyone treating a world-model demo as something a robot or a game can train on this week, and anyone waiting on an API rather than a video.

Try this

Treat Genie 4 as a preview. Do not plan a product on it until it is in the API you can call. If you use the demo, remember DeepMind's own caveat: details drift as the session gets longer.

What we do not know

The post does not print a frame rate or a resolution.
The post does not say how long a session stays coherent.
The public API date is 'coming soon,' not a day.
Object permanence drifts. DeepMind says this gets worse as a session runs longer.

Evidence

Primary Research post. Availability and the role of Gemini 3.8 Flash are from DeepMind's October 27 post. The post does not print a frame rate or a resolution.

Independent This is DeepMind's account of its own preview. No outside lab has published a measurement.

The record

What happened

On October 27, 2026, Google DeepMind posted Genie 4. A person can start a scene and change it while it runs, with language, including the camera, the weather, and new characters. The model generates ambient sound and spoken dialogue with the pictures. DeepMind says this is its first real-time world model to generate audio natively. Gemini Omni is used for appearance and behavior. Gemini 3.8 Flash sits in the loop: it watches the frames, breaks a goal into steps, and checks that each step succeeded. The rollout is a research preview for trusted testers. DeepMind says a wider release will start with Google AI Ultra subscribers in Project Genie, and with early-access partners on the Gemini API and in AI Studio. A waitlist is open. DeepMind says visual details and object permanence can drift as a session runs longer.

Why it matters

A world model you can steer, with sound, is a different tool from a video clip. The limit that matters today is access. Trusted testers have it. A developer does not, and the post gives no frame rate, no resolution, and no session length to set next to the previous Genie.

DeepMind says testers in journalism and games are turning a still photograph or a prompt into a walkable scene. Those are demonstrations, not a measured success rate.

The post says Genie 4 includes the same kind of safety filters DeepMind uses on its other models, and that it was checked internally before the tester preview. It does not publish those test results.

The announcement page on the Google blog is the shorter version of the same October 27 post. It also names trusted testers and a later Ultra and API release.