Genie 3

Genie 3 is Google DeepMind’s AI model for generating environments that react as a person moves through them. Introduced on 5 August 2025, it generates views at 720p and 24 frames per second. A world model produces the next view in response to an action, so turning or moving changes what appears on screen.

The model generates a world as you explore

Google’s original announcement describes interactive environments that retain visual consistency for several minutes. This lets a user move through a generated place and revisit parts of it. The result is a stream of generated views; performance depends on how well the model maintains the scene as actions change.

Project Genie gives subscribers a research prototype

Project Genie lets eligible Google AI Ultra subscribers create and explore generated worlds. Its current product page accepts text, images and Google Street View as starting points. Users can choose a character and perspective, refine the starting image and then explore.

The product is an early research prototype. Check its current eligibility and subscription terms for the region and account being used. A consumer prototype, a research model and a developer service each have their own access conditions.

Waymo built a specialised driving simulator

Waymo’s World Model builds on Genie 3 for autonomous-driving simulation. Its specialised system generates camera and lidar views, including unusual driving scenarios. That industrial adaptation has its own capabilities and testing requirements.

Genie 4 is the newer model

Google announced Genie 4 on 27 October 2026, adding native audio and controls through keyboard input and text. It begins with trusted testers, with further product and partner access planned. The Genie 4 reference follows that newer release and its rollout terms. Genie 3 retains its own original measurements and the history of products built on it.