Google DeepMind released Genie 4 to trusted testers as a research preview, supporting synchronized audio-video generation and two-minute simulations.

Genie 4 creates explorable virtual environments from text prompts and supports keyboard navigation and real-time interaction. Its synchronized sound and image generation includes ambient noise, music, and dialogue, while the reported simulation limit is two minutes.

Access initially covers a small group of trusted testers under a research preview. Google reportedly plans access for Google AI Ultra subscribers and selected developers in the following weeks, but that planned expansion is distinct from current availability.

In the described interactive system, Gemini monitors incoming frames, interprets user intent, decomposes it into steps, and checks whether each step occurred. The team experimented with Argon, while the current real-time experience runs on the lighter Gemini 3.8 Flash model.

Google emphasized that Genie 4 remains experimental: visual details and objects can drift as simulations continue, and reaching two minutes does not guarantee consistency throughout. The company is exploring applications in games, filmmaking, education, and robotics training, while product manager Diego Rivas described the work as research rather than necessarily a finished product.