Google DeepMind unveils Genie 4 world model with native audio

Genie has moved from a research demo to a platform in under three years. The first Genie, announced in February 2024, generated playable 2D worlds from an image. By August 2025, Genie 3 could be walked around in real time at 720p and 24fps, holding visual memory for about a minute. That progression explains why Genie 4's native audio and dialogue are a bigger step than a prettier picture: the model now produces a world you can hear and address, not just watch.

The architectural bet is a division of labor. Genie 4 renders pixels and sound, while a Gemini model acts as an in-loop guide that watches the scene, breaks goals into steps and decides whether each step succeeded. Google DeepMind calls this a first step toward a "hybrid world model" that combines language-model reasoning with pixel-level representation. If it works, the same setup could let a language model direct a simulated environment for robot training or scenario testing without hand-coding every scene.

What the announcement leaves open is how much that costs. Google has not published resolution, fps, generation time, latency or pricing for Genie 4, so the 720p, 24fps and few-minute figures from Genie 3 should not be assumed to carry over. For teams planning work, the practical signals will be the API's price and latency, the first independent reports of long-session behavior, and whether the promised early access actually opens beyond trusted testers.