Google DeepMind announced Genie 4 on Oct. 27, 2026, presenting the world model as a way to create explorable, responsive AI-generated environments with sound. It has begun rolling out a research preview to trusted testers, while broader access remains planned.
In DeepMind’s description of the model, users navigate with keyboard controls and can type instructions as a scene unfolds, such as changing the weather, moving the camera or adding a person. The company calls Genie 4 its first real-time world model to generate audio natively, including ambient sounds and dialogue that adapts to user input or events. It also claims more realistic simulations of aerodynamics and collisions, along with characters that move more naturally.
DeepMind says Gemini 3.8 Flash helps guide a simulation by observing the scene, breaking goals into steps and reviewing incoming frames. The company says Gemini Omni helps anchor how the simulated world looks and behaves. DeepMind says testers are exploring uses including film pre-visualization, historical-scene exploration and game-prototype design; the company also describes robot and AI-agent training environments as a potential use.
Who can try it, and what can go wrong
As of its announcement, DeepMind said the research preview had begun rolling out to trusted testers. Access for Google AI Ultra subscribers through Project Genie and for early developer partners through the Gemini API and Google AI Studio was still to come. The company offered a Project Genie waitlist but gave no release date for those groups.
DeepMind calls Genie 4 an early, experimental model. It warns that visual details and object permanence can drift during longer simulations, after complex edits or while navigating a scene. Audio accuracy, legible text, spatial accuracy and long-term consistency remain areas of research, according to the company.
Comments