Google's Genie 4 adds sound and talking people to live generated scenes, but only invited testers can use it
Google DeepMind's Genie 4 draws a walkable scene live, now with sound and people who speak, and Google pitches it for planning shots. Only invited testers have it, with no date, no developer pricing and no word on saving footage.
Google DeepMind announced Genie 4 on October 27. Google calls it a world model: instead of handing you a finished clip, it draws a scene live as you move through it. You walk and turn the camera with your keyboard, and type instructions to change the weather, add a person or set something in motion.
What is new is sound and people. Every scene now carries its own audio, from wind and water to music, and the people in it move, react and speak, answering what you type or what happens around them. Google says one of its Gemini models watches the scene as it runs so it holds together. The Verge reports that a scene now holds together for about two minutes, up from one minute on Genie 3.
The film pitch is direct. Google says filmmakers among its testers use Genie 4 as a "draft mode" for planning shots, stepping into a scene and running several takes while they change camera angles, lighting, dialogue and action. The two riders above come from Google's own launch film, and neither of them is a real person.
The catch is access and output. Genie 4 is a research preview for invited testers only. Google says it is coming soon to its AI Ultra subscribers in Project Genie, a plan Ars Technica says starts at $100 a month, and to early partners on its developer platform. There is no date and no developer pricing yet. Google has not said whether you can save footage or camera moves, and it warns that details drift the longer a scene runs. Every frame shown so far is Google's own demo.
Our take: for planning a shot, a scene you can walk through, light and talk to beats a clip you can only watch. Until filmmakers can take footage out of it, Genie 4 is a rehearsal room, not a camera.
Source: Google DeepMind blog, Oct 27, 2026, with reporting from The Verge and Ars Technica.
