Google's Genie 4 builds worlds you can walk into and talk to, but only testers can try it

Picture a filmmaker who wants to test a scene before she hires a crew. She types a few lines, steps into a busy street, hears the traffic, and asks a passer-by a question. The passer-by answers. That is the promise Google DeepMind made on 27 October with Genie 4, its newest world model.

What is a world model in AI? It is a system that predicts what happens next in a scene you can move around in, like a video game built on the fly. Genie 4 adds sound, talking characters and live changes by text. For now, though, only a small group of invited testers can use it.

The big AI labs are racing to build these simulators. TechCrunch notes that video startup Runway released its second world model a couple of months ago. The bet is that AI which understands how the world moves will be useful in games, films, classrooms and, above all, robot training. Google already puts Gemini in everyday tools like the Gmail AI Inbox; this is a longer bet.

What Genie 4 adds: sound, live edits and steadier scenes

• You steer it. You move with normal keyboard keys. Type a change, such as rain or a new person, and the scene adapts while it runs.
• It makes sound. Background noise, music and characters who speak are created together with the picture. DeepMind calls it its first real-time world model to do this.
• More natural movement. Collisions and air flow look more real, and people move and react more like people.
• Gemini acts as director. Google says Gemini 3.8 Flash watches the scene, breaks a goal into steps and checks each step worked.
• Longer scenes. A world holds together for about two minutes, up from about one minute for Genie 3, according to The Verge and TechCrunch, citing Google.

In a research test, Gemini 4 Argon planned the moves while Genie 4 simulated Apptronik’s Apollo humanoid robot playing checkers.

Two minutes of wonder, then the details drift

Google is open about the limits. In its own words, visual details and object permanence “can still shift or drift over time”, especially in long sessions or after many edits. Accurate audio, readable text and complex physics are still being researched. Diego Rivas, a Google DeepMind engineer, told reporters it is still early days and the final product is not yet clear.

There is a harder question too. Lifelike talking people, and scenes rebuilt from one archival photo, could make fake footage more convincing. Google says all Genie 4 video and audio carries SynthID, its invisible watermark, and that it ran safety tests before the preview. A watermark helps only if platforms and newsrooms check for it.

Google has not shared a price, a list of countries or a firm date.

When you can try Google Genie 4, and who should care

For now, Genie 4 is a research preview for trusted testers. Google says it is coming soon to Google AI Ultra subscribers, its top paid plan, inside Project Genie, its app for making interactive worlds. It is also coming to early-access developers on the Gemini API and Google AI Studio; TechCrunch reports “in the coming weeks”. A waitlist is open.

If you already pay for AI Ultra, the waitlist is your way in. If you teach, make games or plan film shoots, watch what the testers report; Google says they come from creative work, journalism, education, gaming and engineering. If you mostly watch videos online, the habit that realistic AI voices like Block’s Dessa already call for applies here too: pause before trusting a very real-looking clip, and look for AI labels.

Why Genie 4 matters more for robots than for gamers, for now

There is a lot to welcome here. Google is clear about what does not work yet, it watermarks everything, and it is starting small instead of rushing to millions of users. In our analysis, letting robots practise safely and cheaply in simulated worlds looks closer than a full game or film made by AI.

The weaker side is just as real. Two minutes is short. Access starts with the most expensive plan. And lifelike talking people raise worries about fakes that a watermark alone will not settle.

The likelier outcome is steady progress, not a sudden change in how games or films are made. On balance, Genie 4 looks like a meaningful step for robot training and creative drafts, and its careful, watermarked rollout is the right approach, but for most people it is a promise to watch rather than a tool to use.

The real test is not the demo reel. It is what testers build with it, and whether those two minutes can stretch into something a teacher or a small studio can rely on.