Models · 2026-10-27

Google unveils Genie 4: AI world model now with audio and more realistic people

Google DeepMind is preparing Genie 4, a new version of its model that creates interactive virtual worlds from text descriptions, with access at first limited to paid subscriptions and a waitlist.

Models

What Genie 4 is

Genie 4 is a "world model" from Google DeepMind. It creates a virtual environment you can explore just by describing what you want to see. You don't need to pre-build locations or program objects: the model generates the scene frame by frame in real time.

Its predecessor, Genie 3, caused concern in the gaming industry last year. Google, however, says such models are about more than games. According to the company, world models are a key step toward so-called artificial general intelligence (AGI).

What's new

Genie 4 keeps scenes closer to what you asked for. It uses Gemini 3.8 Flash as a guiding agent that breaks down goals and checks each frame. There is a limit, though: a world stays consistent for only a couple of minutes before it can forget how things started.

You can change the weather, add elements, or move the camera during exploration. Physics is more realistic, with better aerodynamics and collisions. People in scenes should look more natural and less uncanny.

Genie 4 is the first DeepMind world model with native audio. You can add ambient sounds, and eventually human speech that adapts as the scene changes. The company promises real interactive dialog, but for now that is a promise.

Who gets access and when

Developers will be able to use Genie 4 through the Gemini API and AI Studio, but pricing has not been announced. Consumers will first get access in the most expensive AI Ultra plan, which starts at $100 per month. Google may later expand access to AI Pro and AI Plus plans. Access will be handled through Project Genie, which limits what users can do.

No specific launch date has been set. To get early access, you can join the waitlist at the link in this article.

Useful from the source

Sources: 🔗 Ars Technica — AI

Tags