Google has launched Project Genie, a tool based on its Genie 3 AI model that allows users to generate and explore interactive virtual environments from text prompts or images. Available only to subscribers of its premium AI Ultra plan, the system marks the first public access to this advanced world model outside of internal testing. It offers modes like world sketching and remixing, though limited to short 60-second sessions.
Last year, Google introduced Genie 3, an AI world model capable of creating dynamic environments from simple text prompts, with improved long-term memory allowing it to retain details for up to a couple of minutes. Initially restricted to a small group of trusted testers, the model was positioned by Google DeepMind as a tool for training AI agents. Now, in a wider release dubbed Project Genie, it integrates with updated models like Nano Banana Pro and Gemini 3, enabling users to build custom worlds.
To use Project Genie, subscribers must have the AI Ultra plan, priced at $250 per month, and be 18 or older residing in the US. The tool operates through a dedicated web app, separate from the Gemini app or website. Users can start with pre-built worlds or create new ones via 'world sketching,' where Nano Banana Pro first generates a reference image based on a prompt describing the environment, character, camera perspective—such as first-person, third-person, or isometric—and exploration style. Adjustments can be made to the sketch before Genie 3 generates the interactive video.
The resulting simulation renders at 720p resolution and 24 frames per second, allowing navigation with WASD keys in near real-time. Explorations are capped at 60 seconds, after which users can rerun prompts for varied outcomes or remix existing worlds with new characters and styles. Videos of sessions are downloadable. Google emphasizes that Project Genie remains a research prototype, with limitations including input lag, occasional inaccuracies in physics modeling, and evolving content restrictions—for instance, prompts mimicking Nintendo games like Super Mario were initially allowed but later blocked to respect third-party interests.
While not a full game engine, the system simulates physical interactions without traditional mechanics. Google plans to expand access beyond the premium tier over time.
