Google DeepMind announced Genie 4 on October 27 as an experimental model for interactive simulated worlds. The company says it can generate visual scenes and accompanying sound together, including ambient noise, music and dialogue that responds as a simulated scene changes.
Users can navigate with keyboard controls and steer events through natural-language instructions. Demonstrated changes include moving the camera, altering weather and adding people or objects. Google's account describes Gemini models helping maintain the scene's context while Genie produces the evolving visual and sound experience.
One research arrangement uses Gemini 3.8 Flash to observe incoming frames, divide a goal into steps and check progress. DeepMind also described experiments connecting Genie with Gemini 4 Argon and robotics research. These are simulated demonstrations, not evidence that a physical robot can reliably perform every depicted task.
Trusted testers are exploring filmmaking previews, interactive game concepts, historical environments and engineering scenarios. Such uses can help teams examine an idea before building it. A convincing simulation, however, should not be treated as a verified reconstruction of a historical event or a validated engineering result.
DeepMind explicitly says visual details and object permanence can drift, particularly during longer sessions or repeated complex edits. It also lists physical interactions, audio accuracy, readable text, spatial accuracy and long-term consistency as unfinished research areas. The preview's limitations remain part of the announcement.
Access initially went to trusted testers. Google says Project Genie access for AI Ultra subscribers and early access for Gemini API and AI Studio partners are coming later; it has opened a waitlist. The company also says generated audio and video will carry SynthID watermarking to help identify AI-made content.