Google DeepMind unveils Genie 4 world model with generated audio, dialogue and Gemini-steered events

Genie 4 is DeepMind's biggest world-model jump yet. It renders explorable environments frame by frame in real time with user steering, and adds native audio for the first time: ambient sound, effects, speech and character dialogue generated in sync with the visuals. Google says environments stay coherent for up to two minutes, pairing Gemini Omni for visual consistency with Gemini reasoning models that plan multistep events and adapt the world to the user.

The Decoder framed the architecture as 'a modern spin on agent designs that predate LLMs': a learned simulator with a planner on top, close to classic model-based RL. That is why DeepMind pitches Genie for agent training environments as much as for entertainment. Bloomberg noted earlier versions rattled video game stocks, and this upgrade renews that question.

Availability is staged. Trusted testers and Project Genie users get it now. Google AI Ultra subscribers follow in the coming weeks, along with early-access partners via the Gemini API and AI Studio.

The skeptics are specific. On Hacker News, developers said scene elements morph or hallucinate once they leave the field of view, meaning there is no true object permanence. Roboticists and graphics engineers argued pure autoregressive frame generation cannot replace explicit 3D representations and drifts over multi-minute sessions. Two minutes of coherence is impressive for demos, but it is short for training long-horizon agents. Watch for API pricing and whether developers can export state for reproducible simulation.