Google DeepMind announced Genie 4 on October 27, the next version of its interactive world model. Like its predecessors, Genie 4 turns a text prompt or an image into an environment a user can move through in real time with keyboard controls. The new release aims at three weaknesses of earlier versions: realism, sound and the ability to change a scene once it is running.
The most visible change is audio. DeepMind says Genie 4 is its first real-time world model that produces sound natively, so each scene arrives with its own ambient noise, from wind and water to music played by characters on screen. It can also generate spoken dialogue, with characters that talk and react to what the user does or to events in the simulated world.
Users can now rewrite a scene while it plays. A typed instruction can change the weather, move the camera, or drop a new person or object into the world, and Gemini reinterprets the request so it fits what is already on screen. Between user inputs, Gemini keeps watching the scene and nudging it so the world continues to make sense. DeepMind also claims better handling of physics such as collisions and airflow, and much more natural human movement.
Under the hood, Genie 4 borrows from the Gemini family. Gemini Omni grounds how things look and behave, while Gemini 3.8 Flash acts as a planner in the loop: it splits a goal into steps, checks incoming frames and confirms each step before moving on. In a robotics experiment, Gemini 4 Argon fed step-by-step captions to Genie 4 to simulate Apptronik's Apollo humanoid playing checkers, which DeepMind presents as a way to rehearse robot tasks in imagined settings before trying them for real.
According to TechCrunch and The Verge, a Genie 4 scene can stay coherent for about two minutes, roughly twice the limit of Genie 3, which Google previewed in August 2025. TechCrunch places the launch in a crowded field that includes Runway's GWM Worlds 2 and a wave of consolidation among world-model startups. Google DeepMind researchers told reporters that hybrid systems, in which language models plan and Genie renders, are where they expect the technology to go.
Access is narrow for now. Only trusted testers can use Genie 4 today. Google says paying Google AI Ultra users will get it next through Project Genie, along with selected developers using the Gemini API and Google AI Studio, and it has opened a waitlist. Testers in film, journalism, education, gaming and engineering have used it for storyboarding, recreating archival photographs as explorable scenes and prototyping game ideas. Every video and audio output carries Google's SynthID watermark, and DeepMind says the model went through internal safety evaluations before the preview.
DeepMind is candid that Genie 4 remains experimental. Object permanence and fine detail can drift during long sessions or after many edits, and physical interactions, audio accuracy, readable text and spatial precision all still need work. Neither TechCrunch nor The Verge had used the model themselves, so the quality shown in Google's demo clips has not yet been tested independently.