Google started rolling out access to Project Genie, the public interface for its Genie 3 world model. The catch: you need a Google AI Ultra subscription, you need to be in the US, and your explorations cap out at 60 seconds.
Genie 3 differs from video generators like Veo in a fundamental way. It doesn't produce a fixed clip. Instead, it renders environments frame-by-frame at 20-24 fps as you navigate, maintaining consistency when you revisit areas for up to a minute. The model outputs at 720p, which Google positions as sufficient for training AI agents on real-world complexities.
The workflow uses Google's Nano Banana Pro image model to sketch a preview before you enter. You describe your world and character separately, then explore via keyboard or controller. A library of pre-built worlds is available for remixing, and sessions can be downloaded as video.
Google's model page acknowledges the limitations openly: worlds don't perfectly replicate real locations, text rendering is unreliable, and promptable events announced in August aren't included yet. International availability is promised but undated.
The Bottom Line: Google is treating this as an AGI research milestone, not a consumer product, using Ultra subscribers as a scaled testing cohort for its world simulation work.
QUICK FACTS
- Resolution: 720p at 20-24 fps
- Session length: 60 seconds maximum (company-reported)
- Availability: US only, Google AI Ultra subscribers (18+)
- Preview model: Nano Banana Pro
- World consistency: Up to 1 minute of recall
