We’re announcing Genie 4, our new world model for real-time interactive simulation.

What it claims

Genie 4 is a Google DeepMind world model for real-time interactive simulation, announced as research. Users start an event and then steer the scene in real time with natural-language commands. The post gives three examples: moving the camera, changing the weather and introducing new characters. Each world also generates audio on the fly: ambient sound, music and interactive dialogue. Google DeepMind presents world simulation as a step toward AGI. It names two uses: safe training environments for robots and agents, and creative tools for filmmaking, journalism, gaming and manufacturing.

How it works

The post describes the design only at the level of which Gemini models contribute. Gemini Omni anchors simulations in real-world appearance and behavior. Unnamed "other Gemini models" supply the reasoning that keeps a simulation coherent and continuous over time. The post does not describe the architecture, training data, model size, frame rate, resolution or session length.

How to use it

The rollout has three stages, and the API stage is gated:

1. A limited group of trusted testers has access now.
2. Google AI Ultra subscribers get access "soon" through Project Genie, a consumer product.
3. In the Gemini API and AI Studio, access is limited to "early access partners". General developer access is not announced.

The post gives no API shape, pricing, rate limits, model identifier or release date. Demos are on the Google DeepMind blog, not in this post.

Limits

The post is a short announcement. It has no benchmarks, no comparison with Genie 3 or any other world model, and no stated failure modes. The claim of "unprecedented realism" has no supporting measurement. The post does not show that Genie 4 works as a robot or agent training environment. It names that use as a goal.