Genie 3

Genie 3 generates its environments frame by frame, in real time, at photorealistic quality, for up to one minute. Google DeepMind announced it in August 2025 and shipped it in February 2026 inside Project Genie, a web application open to Google US Ultra subscribers. The lineage is short and steep: Genie 1 was a research paper, Genie 2 (December 2024) produced ten seconds at low resolution with no real-time control, and Genie 3 replaced it.

The technical problem is harder than standard video generation. A video model can retroactively adjust earlier frames to keep a sequence coherent; Genie 3 cannot. Each frame has to stay consistent with the past and with the user's immediate keyboard or mouse action, without any knowledge of the inputs still to come. The physics is not preprogrammed. The model learned gravity, collisions and movement dynamics by observing large volumes of 2D platformer game footage, an emergent understanding rather than a rule engine.

The one-minute ceiling is a compromise. World dynamism tends to decay over longer runs, and serving costs stay high. Google DeepMind also acknowledges a complexity ceiling on game mechanics, imperfect control fidelity, and training-data bias.

Against OpenAI Sora and Runway Gen-3, interactive control is the differentiator. The heavier consequence sits elsewhere: Simmer already uses Genie 3 to train agents to accomplish goals in arbitrary 3D worlds, which points the model at embodied intelligence rather than entertainment. The team puts itself at 50% of its stated vision.