Published: 28 October 2026
Google DeepMind presented Genie 4 on 27 October: a world model that turns a text prompt into a scene you walk through with your keyboard, complete with sound and characters who talk back. Gemini models plan what happens in the scene and check each step against the picture. For now only trusted testers can use it, with Google AI Ultra subscribers and selected developers next in line.
You steer the scene by typing, and Gemini rewrites your request to fit it
A world model is an AI that simulates a place frame by frame and reacts to what you do, instead of rendering a finished video. In Genie 4 you move and turn the camera with standard keyboard controls. During a session you can type a change, such as rain, a new camera angle or an extra person. Gemini then adapts that prompt to the running scene.
The visuals come from Gemini Omni, Google’s video model, which handles how people and places look and move. On top of that, Gemini 3.8 Flash acts as a guide. It watches the scene, splits a goal into steps, checks new frames to confirm each step and plans again when you step in.
Sound is generated in the same loop: wind, waves, music and dialogue that adapts to events. Google calls Genie 4 the first real-time world model that creates audio natively. TechCrunch points out that Runway’s GWM Worlds 2, released a couple of months earlier, is also playable and also generates audio. The word first is therefore Google’s claim, not an established fact.
Apptronik’s Apollo robot played checkers without a real robot in the room
The most telling demo has nothing to do with games. Gemini 4 Argon split a game of checkers into steps, and Genie 4 rendered Apptronik’s Apollo humanoid robot carrying them out in real time. No physical robot took part. Google presents this as a way to build training and test environments for robots and AI agents, next to uses in film previews, journalism, education and game prototyping.
For organisations, that robot demo is the part to watch. Think of situations that are rare, costly or unsafe to stage, such as a machine fault or a near miss in a warehouse. A robot or agent could rehearse those in simulation before it meets them for real. Today that is a research direction, not a product. The lesson you can apply now sits in Gemini’s role: break the goal into steps, check the result of each step and plan again when the situation changes. That same loop is what keeps business agents reliable when they handle documents or run a workflow, so build it into any agent you deploy.
Google gave Genie 3 a resolution and frame rate, and Genie 4 none
When Google showed Genie 3 in August 2025, it stated 720p at 24 frames per second. For Genie 4 it publishes no resolution, frame rate, delay or test scores, and no comparison with Genie 3 beyond demo clips. TechCrunch reports that a Genie 4 scene stays consistent for up to two minutes, double the previous limit. Google’s own model page only speaks of “a few minutes of continuous interaction”. In Project Genie, the current Genie 3 version gives you 60 seconds to explore each world.
Google is open about the weak spots. Visual details drift and objects can vanish in longer sessions, after several complex prompt edits or while you move around. Complex physics, accurate real-time audio, readable text and spatial accuracy are still research problems. Every video and audio track carries Google’s SynthID watermark. All claims so far come from Google itself, and no independent test is available yet.
How do you get access to Genie 4?
For now only through a waitlist, because Genie 4 runs as a research preview for trusted testers. Google says AI Ultra subscribers will get it in Project Genie, its web app for building worlds, and TechCrunch puts that in the coming weeks. Early access partners of the Gemini API and Google AI Studio follow as well. Google has not announced a price or a general release.
Google names no countries for Genie 4. Project Genie itself has been rolling out since May to eligible AI Ultra subscribers worldwide aged 18 and over. If Genie 4 follows that route, Ultra subscribers in the Netherlands and the rest of the EU would get it there too, but Google has not confirmed this. One existing feature is already limited by country: building a world from real places with Street View works only for US locations.
Most organisations can let Genie 4 pass for now: there is nothing to buy, no API price and no measurement to judge it by. Teams that build robots, train AI agents in visual environments or make film and training material have a reason to join the waitlist. Go in with one concrete scenario that is hard to stage in real life. Hold off on plans until Google publishes API access, a price and figures on how long a world stays consistent.
For more information, visit the official announcement of Genie 4 on the Google DeepMind blog.
