New Delhi: Google DeepMind has unveiled Genie 4, the latest iteration of its “world model” technology, that creates interactive environments based on a few words. The announcement was made via a summary of posts on X on October 27, 2026.
The model is said to create photorealistic environments with real-time physics and native audio, as well as responsive characters driven by Gemini models, which are responsive. Users use natural language controls to navigate a scene and control the events and camera perspectives along the way. Settings that were posted online included snowy landscapes and busy city streets.
Early reactions
First testers have already given it their feedback. Justine Moore, a partner at the venture firm a16z, was one of those touting the model. It has the power to create a completely interactive GTA world out of a simple prompt, she said. The Genie 4 is rolling out to trusted users first, and will be available to Google AI Ultra subscribers soon.
This information is based on social media posts and a bot-generated summary of them, and not on a full technical release. Resolutions, frame rates and session durations are not verified independently yet, so they need to be taken with a grain of salt until Google releases more information.
What a world model is
World Model is a set of AI systems that predict the changes of an environment due to actions. Google says it’s akin to a language model predicting the next word, but instead the model predicts what happens next in the world as an agent interacts with it.
In February 2024 the original Genie was revealed, but was a research project only. It was trained on Internet videos and could transform any image into a rudimentary, playable 2D world. It was also slow. It was not in real time, it was running at about a frame per second, and DeepMind did not release the code or weights, researchers said.
A quick progression
Since then, there’s been rapid development. Project Genie was based on Genie 3, and was released in early 2026, producing explorable 3D worlds at 720p-24fps. This feature was restricted to U.S. subscribers on Google’s $250 per month AI Ultra subscription and generated worlds were gone after 60 seconds. At the time, a Google spokesperson stated that it’s not the game engine to build a complete game experience.
These all seem to indicate a great jump: photorealism, built-in sound, characters that react to the player and Gemini behind the scenes.
Why it matters
Google frames world models as a creative tool for designers and artists: use to brainstorm and rapidly prototype. There’s been greater commentary on a deeper motive. DeepMind’s mission is to create Artificial General Intelligence, and an AI capable of predicting the effects of actions on an environment is a step towards reasoning about the physical world. The same technology could be used to simulate training robots and autonomous vehicles.
Google is not the only one in this race. There are several startups, including World Labs, who have raised a lot of money, working on similar technology, and DeepMind has dedicated teams for world models.
What to watch
There are still a number of questions unanswered. How long will a Genie 4 world remain coherent? Who pays and who is paying for access? But how close do the demos come to daily life and not mere showcases? A few snapshots of la(deficiant) quality doesn’t make a game, and so comparisons with GTA have to be taken with a pinch of salt.
At this time, Genie 4 appears to be another indication of the rapid progress of interactive AI worlds from research demos to actual product. The greater exposure for the Ultra subscribers will tell if the demos live outside Google’s own dock.
