Google DeepMind unveils Genie 4 world model with audio and real-time prompt edits

Google DeepMind announced Genie 4, the latest version of its world model, which generates interactive environments from text prompts frame by frame in real time. It launched as a research preview for trusted testers. Google says the model has more realistic worlds, lets users change a scene with a natural-language prompt mid-simulation, generates audio, and uses Gemini 3.8 Flash in its simulations.

Experiences now run about two minutes, up from one minute with Genie 3. Users can change the weather or camera angle, or add people and objects, while the simulation runs. Google says physics such as aerodynamics and collisions are better, and that people move and respond to their surroundings more naturally.

Genie 4 is DeepMind's first world model with native audio. TechCrunch reports it produces music, ambient noise and interactive dialogue that matches the scene. Gemini works as a director, planning steps and checking frames while Genie draws the visuals. In a Gemini Robotics experiment, a Gemini 4 Argon model split tasks into steps that Genie 4 simulated live.

On availability, TechCrunch reports the model will reach Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio in the coming weeks. Ars Technica says Ultra plans start at $100 per month and developer pricing is not yet available. A waitlist is open for Project Genie.

Google positions Genie 4 for pre-visualization in film, game prototyping and robot training. Diego Rivas, a DeepMind engineer, told reporters that real-time generation matters most for ideation and storyboarding. The robotics experiment points to testing scenarios in simulation before running them in the real world, though it remains an experiment.

The release arrives as the field crowds and consolidates. TechCrunch notes Runway's GWM Worlds 2, AMD's pending acquisition of World Labs, Amazon's reported talks to buy Decart, and Rivian's purchase of Third Dimension for self-driving simulation. Genie 3 never moved beyond the limited Project Genie web experience, so the gap between demos and a product remains the question for Genie 4.

The limits are significant. Google says details and objects can drift as simulations run longer, and Ars Technica notes worlds forget how they began after a couple of minutes. The team has not worked out how to guarantee accurate audio or legible in-world text. Reviewers, including The Verge's Jay Peters, have not been able to try it.

Watch for the Project Genie rollout to Ultra subscribers and the opening of Gemini API access, both promised within weeks. Pricing for developers and whether AI Pro and AI Plus tiers get access are open questions.

Rivas said it is not yet clear what the end-to-end product experience will be, and that using Gemini Omni to turn prototypes into creative assets is still a way off.