Genie 3 (Project Genie)
Released Jan 29, 2026
Google DeepMind's general-purpose real-time world model, available to consumers through the Project Genie prototype (with Nano Banana Pro and Gemini) to sketch, explore and remix playable worlds from text and images.
- Output
- Live video stream
- Interactive
- Yes
- Inputs
- text, image, actions
- Max session
- 60 s
- Resolution
- 720p
- Frame rate
- 24 fps
- Access
- Research preview
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Real-time navigable photorealistic worlds at 720p, 20-24 fps
- Environments stay largely consistent for several minutes, with visual memory of about a minute
- Street View-grounded worlds of real U.S. places (since 2026-05-19)
- Available to Google AI Ultra subscribers globally (18+)
Weaknesses
- Project Genie generations capped at 60 seconds
- Requires the $200 Google AI Ultra plan; not available on Ultra for Business; no API
- Google notes higher control latency and imperfect prompt/physics adherence; promptable world events not in the prototype
Genie 3 was announced 2025-08-05 as a limited research preview for select academics/creators; releaseDate is the first public launch via Project Genie (2026-01-29, U.S. Ultra subscribers), expanded to Ultra subscribers globally on 2026-05-19. Google calls Project Genie 'an experimental research prototype', hence access 'research-preview' even though any eligible paying subscriber can use it. fps: DeepMind model page says 20-24 fps, the Genie 3 blog says 24. maxSessionSec 60 = Project Genie cap; the model itself supports 'a few minutes' of continuous interaction. Successor Genie 4 announced 2026-10-27 as a trusted-tester research preview (separate entry). Google says it is coming to Project Genie for Ultra subscribers; until then Project Genie runs Genie 3. The deepmind.google/models/genie page now presents Genie 4.