Google built AI agents a world to practice in. The physics still aren’t quite right.

On Tuesday, Google DeepMind unveiled Genie 4, an advanced world model capable of rendering interactive environments in real-time with simulated physics, human motion, and native audio. This innovative approach provides developers with a platform to test AI agents in multi-step tasks before deployment, according to Tejas Kulkarni, a research director at Google DeepMind.

The technology combines Gemini models to maintain the simulation, with Gemini Omni shaping the environment, Gemini 3.8 Flash planning the next steps, and Genie 4 rendering the results in real-time. Users can interact with the world using a keyboard, alter the camera angle, or input text prompts to modify the environment without restarting the simulation.

Google has demonstrated how Genie 4 could aid its Gemini Robotics research, allowing developers to test agents in changing environments without rebuilding simulations. However, the physics within Genie 4 is not always consistent, with objects occasionally shifting position or behaving unpredictably during extended sessions. This inconsistency complicates agent evaluations, particularly when tasks require remembering locations or returning to earlier states.

While game developers and enterprise partners are exploring the potential of Genie 4 for prototyping and generating testing scenarios, the company has not published benchmarks showing the technology's performance across repeated agent tests. Currently, only trusted testers have access to Genie 4, with plans to expand availability to Google AI Ultra subscribers through Project Genie and select partners via the Gemini API and Google AI Studio. Pricing, a wider release, and performance guarantees have not been disclosed.