Google puts Genie 3 behind a paywall with the Project Genie prototype

From trusted testers to a paying subscriber tier

In August, Google previewed Genie 3 as a general-purpose world model available to trusted testers. Project Genie is the next distribution step: a web app that packages that model for Google AI Ultra subscribers in the U.S. who are 18 or older. Access begins rolling out on January 29, with expansion to more territories described only as coming 'in due course.'

The choice of gate matters. Rather than a broad public launch or a continued closed test, Google routed the prototype through its most expensive subscription tier. That keeps the audience small and self-selected while still moving the technology out of the pure research setting and into the hands of people who will use it for their own projects.

Three capabilities stitched from three models

Project Genie is not Genie 3 alone. It is a prototype powered by three components working together: Genie 3 for the world simulation, Nano Banana Pro for image work, and Gemini. The product surface is organized into three actions — world sketching, world exploration, and world remixing.

World sketching uses Nano Banana Pro to let users preview and fine-tune an environment as an image before entering it, and to set perspective such as first- or third-person. Exploration is where Genie 3's defining behavior shows up: as the announcement puts it, the model 'generates the path ahead in real time based on the actions you take,' rather than loading a pre-built static scene. Remixing lets users build on existing prompts, pull from a curated gallery, or use a randomizer, and download video of their explorations.

Unlike explorable experiences in static 3D snapshots, Genie 3 generates the path ahead in real time as you move and interact with the world.Montana Labs

The limitations are the most useful part of the announcement

Google is unusually direct about what does not work yet. Generated worlds may not look true-to-life or follow prompts, images, or real-world physics. Characters can be less controllable and can suffer higher control latency. Generations are capped at 60 seconds. And notably, promptable events — the ability to change the world as you explore it, a capability shown in August — are not included in this prototype at all.

That last point is worth flagging: the shipped product is a subset of the August research demo, not a superset. The 60-second ceiling and the missing event system tell you this is a controlled slice chosen for stability, not a full expression of the model's demonstrated range.

What a paywalled world model signals for applied builders

For teams watching world models, the concrete signal here is that real-time interactive generation has crossed from lab demo into a product with a price, a rating restriction, a time limit, and a documented failure list. The physics-and-consistency claims Google makes — spanning robotics, animation, and historical settings — are still framed against a model that its own team calls early.

The honest read is that Project Genie is a data-gathering exercise dressed as a product. Google states it wants to 'better understand how people will use world models' across AI research and generative media. Anyone evaluating this technology should treat the 60-second cap, control latency, and absent promptable events not as footnotes but as the actual boundary of what is production-viable today.