Google DeepMind has introduced Genie 4, which Ars Technica says adds native audio, human speech and more realistic physics to interactive AI-generated worlds, with access opening first to developers and AI Ultra subscribers.
Google DeepMind has unveiled Genie 4, the latest version of its Genie world model, according to Ars Technica, and a post titled "Genie 4: A new generation of world simulation" appears on DeepMind's blog under Models for October 2026. A world model generates an interactive virtual environment frame by frame in real time from a text prompt, without pre-rendered scenes or hand-programmed interactions. Ars Technica reports that Google regards such models as a step towards artificial general intelligence, and that DeepMind says they could supply endless training environments for robots or speed up prototyping in creative fields.
According to Ars Technica, Genie 4 uses a Gemini model, Gemini 3.8 Flash, as a guiding agent that breaks down goals and evaluates each generated frame, and users can still change conditions such as the weather or camera midstream. DeepMind says the model handles aerodynamics and collisions better and renders people with more natural movement. It is described as the first DeepMind world model with native audio, including ambient sound and human speech, with dialogue that is meant to adapt as the simulation changes. These are the company's claims and have not been independently tested, and DeepMind acknowledges that Genie 4 worlds stay consistent for only a couple of minutes before losing track of how they began.
No release date has been given. Ars Technica reports that a small group of trusted testers is already working with the model, that developers are due to get it through the Gemini API and AI Studio at prices not yet announced, and that consumers will be offered access in phases through Project Genie, starting with Google's most expensive AI Ultra plans, which begin at $100 a month. Google may extend access to cheaper plans later, and a waitlist is open. DeepMind stresses that world models remain an active area of research, noting that Genie 3 never grew beyond the limited Project Genie web experience.
