Google DeepMind’s Genie 4 generates playable game worlds in real time, with physics and sound built in

Movement through these worlds works with ordinary keyboard controls while the environment keeps generating itself around the user. Typing a plain-language prompt mid-session can shift the weather, reposition the camera, or drop new people and objects into the scene, and Gemini adjusts each change to fit what’s already there rather than resetting the world from scratch. Characters inside these spaces move and react to their surroundings, can be redirected on the fly, and hold conversations that shift based on what the user says or does.

Under the hood, Gemini Omni keeps the simulation grounded in how things actually look and behave, other Gemini models handle reasoning and continuity between prompts, and a separate Gemini model acts as a kind of director, planning out steps, checking whether the scene matches what was asked for, and course-correcting when the user intervenes. The model also simulates physics like aerodynamics and object collisions, and generates ambient sound suited to each setting, wind through a canyon, waves against a shore, rather than looping a generic soundtrack over the top.

DeepMind is positioning Genie 4 well beyond entertainment: training environments for robots and AI agents, previsualization for filmmakers, turning old archival photos into walkable spaces, game prototyping, and building out enterprise training scenarios all get a mention as intended uses. For now it’s a research preview limited to trusted testers, with a waitlist open for a wider rollout to Google AI Ultra subscribers through Project Genie, plus early access for select partners through the Gemini API and Google AI Studio.

DeepMind is upfront about where the model still struggles. Visual details and object permanence can drift during longer sessions or after a string of edits, and the company lists complex physical interactions, audio accuracy, legible in-world text, spatial accuracy, and long-term consistency as areas still under active research. Every generated clip carries a SynthID watermark, and the system went through internal safety testing before reaching trusted testers.