Google Opened Its AI World-Generator to the Public, and Each World Lasts 60 Seconds

Type a sentence describing a place, and within seconds you can walk around inside it. That is the pitch behind Project Genie, and as of late January 2026, Google has let ordinary subscribers try it for themselves, sixty seconds at a time.

60 seconds
The hard limit on how long a generated world can be explored per session, a restriction Google says comes from the dedicated computing power each session requires, not from the underlying model itself.

Why does this matter?

Most generative AI so far has produced things people watch: text, images, video. A world model instead tries to simulate an environment you can move through and act inside, predicting how a scene should look next based on where you choose to go. Google DeepMind and its rivals see this as a meaningful step toward AI systems that understand physical space, useful for training robots in simulation before they touch the real world, and, more immediately, for building game worlds without hand-crafting every asset.

What does Project Genie actually do?

The prototype chains together three of Google's AI systems. A user first describes an environment and a character in text; Google's Nano Banana Pro image model turns that into a starting image, which the user can adjust. Genie 3, the underlying world model DeepMind first previewed as a research system in August 2025, then generates an explorable 3D space from that image, rendered in real time as the user moves through it with keyboard controls.

Early users found the tool strongest with stylized, fantastical prompts, claymation castles, watercolor forests, dreamlike landscapes, and noticeably weaker at photorealism, where scenes tend to resemble older video games rather than convincing real-world scans. TechCrunch's hands-on account described building a marshmallow-and-chocolate castle world, alongside frequent glitches like characters clipping through walls and controls that occasionally sent them careening in the wrong direction.

Why only 60 seconds?

Genie 3 generates each frame based on the ones before it, which means every active session needs dedicated processing power rather than shared, cached results. Google has said extending sessions much further would either be prohibitively expensive to run at scale or would sharply limit how many people could use the tool at once, so the company chose short sessions over a longer wait list.

What happens next?

Google is not alone in this race. Fei-Fei Li's startup World Labs released a commercial world-model product called Marble in late 2025, AI video company Runway has shipped its own world model, and former Meta chief AI scientist Yann LeCun has started a new venture built around the same idea. For Google, every world a user generates in Project Genie also feeds back into DeepMind's training data, meaning the rough, glitchy public release is doing double duty as both a product test and a data-gathering exercise.

For now, Project Genie is less a game engine than a very short, occasionally broken preview of one. Whether that is a novelty or the first draft of how game worlds and training simulations get built will depend on how quickly the rough edges get sanded down.