Google DeepMind opens Project Genie to early users in the US

Google DeepMind has begun rolling out Project Genie, an experimental interface that lets people create and navigate AI‑generated interactive worlds, to a limited group of users. The company said access starts for adults who subscribe to Google AI Ultra in the United States from 29 January 2026, with plans to expand later. The release is framed as a research prototype to gather feedback, not a full consumer launch.

What the prototype offers

Project Genie is powered by DeepMind’s Genie 3 world model, working alongside the Gemini family and the image tool Nano Banana Pro, according to the company. It is designed to let users build a setting with a short text description or an image, then enter and move through that space while the model renders the path ahead in real time.

• World sketching, users can combine text with generated or uploaded images to shape a scene, pick a perspective such as first person or third person, and define how they wish to traverse it.
• World exploration, as the player moves, Genie generates what lies ahead in real time and allows camera adjustments during the session.
• World remixing, creators can adapt existing prompts from a gallery or a randomiser, and export short video clips of their runs.

Under the hood

DeepMind describes Genie 3 as a general purpose world model that can turn a brief text prompt into a photorealistic, interactive environment, supporting fluid movement at roughly 20 to 24 frames per second and up to 720p output. The lab positions this as progress on stability and control over longer interactions compared to earlier versions.

How does Project Genie work in practice

The experience starts with a prompt. Users outline a world, optionally refine it with an image preview, select a viewpoint, then enter and explore. The system simulates changes as you move, and sessions can be remixed or shared from a curated library so others can iterate. Google says this setup is meant to showcase the research model’s capabilities while surfacing real world usage data from a broader set of testers.

Limits and responsible use

As an early research release, the company points to several constraints. Generated scenes may not always look realistic or adhere to strict physics, control responsiveness can vary, and generations are limited to about 60 seconds. Features previewed in earlier research, for instance, promptable in‑world events, are not part of the current prototype. DeepMind says it is collecting feedback to improve safety and quality before wider availability.

Why it matters

World models are an active research area at DeepMind, which has long used simulated environments to study learning and decision making. In its public materials, the lab frames such systems as a stepping stone towards more capable agents that can reason and act across diverse tasks, while emphasising responsible development.

Leadership context

Project Genie is being developed within Google DeepMind, led by co‑founder and chief executive Demis Hassabis. At the parent company level, Sundar Pichai is chief executive of Alphabet and Google. Their stewardship signals the strategic priority of building and testing advanced models, including world models, across research and creative workflows.

For creators, the prototype lowers the barrier to prototyping interactive experiences by compressing asset creation and world-building into a short prompt. For researchers, it adds another sandbox for studying how agents perceive, plan and adapt. The company’s decision to start with a limited population and to set clear guardrails around duration and behaviour suggests a cautious approach that could evolve based on early feedback in the coming weeks.