Google DeepMind has begun a limited research preview of Genie 4, a model for interactive, sound-enabled simulated worlds, giving selected testers a way to alter scenes while they unfold.
Participants can navigate environments with a keyboard and mouse, then use text prompts to adjust weather, move the camera or add people and objects. Google says Gemini interprets instructions in the context of the scene and helps the generated world develop coherently, allowing users to combine movement with real-time changes.
Genie 4 generates ambient sounds and music as scenes run, and characters can respond to prompts or events. Google’s examples include a saxophone player beside a misty lake and a Japanese train station set in the 1980s. The company says its demonstration recordings are illustrative and some sequences are shortened.
Google has tested the model with people in creative fields, journalism, education, gaming and engineering. The company cites filmmakers exploring scene ideas, photojournalists working with archival images and game developers prototyping environments as examples of potential uses.
The research also includes a robotics simulation. In one demonstration, Gemini 4 Argon breaks a task into steps while Genie 4 simulates an Apptronik Apollo humanoid playing checkers. Google says realistic simulations could help researchers test AI agents before deploying them in the physical world.
Google says scenes can change during extended exploration or after several complex edits, and objects may fail to remain in place. Accurate audio, complex physical interactions, readable text, spatial accuracy and consistency over time also remain research challenges, the company says. Google reports that it conducted internal safety evaluations and that Genie 4-generated video and audio will carry its SynthID watermark.
Google introduced Genie 3 in 2025, describing it as generating interactive worlds at 24 frames per second and 720p resolution, with consistency lasting several minutes. The company has not announced a calendar date for broader Genie 4 access. It plans to offer the model through Project Genie for Google AI Ultra subscribers and through the Gemini API and Google AI Studio for early-access partners.
