← sources

Genie 4 puts a Gemini agent inside the world model, but nobody outside Google can use it yet

source: reddit (reddit:singularity)published 2026-10-27 20:19ingested 2026-10-28 02:37

retained text

Google DeepMind announced Genie 4 today. You describe a place and it builds a scene you can move through in real time, and you can keep typing while it runs to change the weather, move the camera or add people. Scenes now stay consistent for two minutes, double Genie 3, it makes its own sound, and characters talk back. The interesting change is Gemini 3.8 Flash running inside it, breaking goals into steps and checking every frame to see if each step worked, which is an agent loop in a world that’s being generated around it.

It’s a research preview for trusted testers. Google says Ultra subscribers get it in Project Genie (waitlist) and early access developers get it through the Gemini API over the coming weeks, with no price yet. Google’s own list of what doesn’t work lists that objects drift in long sessions, and text, sound accuracy and complex physics aren’t solved. I wrote up what’s new, why world models matter as training grounds for agents, and who can use it: [https://www.anchorterminal.com/blog/genie-4-world-model](https://www.anchorterminal.com/blog/genie-4-world-model)

document fetched 2026-10-28 00:30 · canonical https://reddit.com/r/singularity/comments/1xh3k9u/genie_4_puts_a_gemini_agent_inside_the_world