Google DeepMind announced Genie 4 on 27 October 2026, a world model that generates interactive scenes with their own sound and lets you change them with plain text while they run. DeepMind's Genie model page says it supports "a few minutes of continuous interaction," and it is a research preview for trusted testers. If you build software, treat Genie 4 as a simulator to try for training, testing and early prototypes. Do not treat it as a reason to rebuild a product your customers already use.
What is Genie 4?
Genie 4 is a model that generates an interactive world you can move through and change in real time. In the Genie 4 announcement on the Google DeepMind blog, Raia Hadsell, Diego Rivas and Hang Qi describe it as "our new world model" and call it "natively multimodal." Visuals, audio and speech are produced together with the world, not added afterwards.
A world model is not a game engine. Nobody wrote the physics as code. The model predicts what the next moment should look and sound like, based on what you do. That is why it can surprise you, and why it can drift.
What is new in this version?
Three things stand out in DeepMind's post.
- Native audio. DeepMind says Genie 4 is "our first real-time world model to generate audio natively." Characters can talk to you and to each other.
- Live steering with language. You can "use natural language to modify a simulation in real-time." Type that it starts raining, and the scene changes while you are in it.
- Normal controls. You move and look around with standard keyboard controls, the way you would in a game.
The Genie model page adds that Genie 4 builds on the Gemini family, with Gemini models handling grounding and reasoning so that the world stays coherent.
Who can use Genie 4 today?
Trusted testers only. DeepMind says Genie 4 is coming soon to Google AI Ultra subscribers in Project Genie, and to early access partners of the Gemini API and Google AI Studio. There is a waitlist. There is no general developer release and no public price.
That matters for planning. You cannot build a customer feature on a model you cannot buy. Anything you do with Genie 4 this quarter is research, and should be budgeted and scoped that way.
Where would a world model actually help?
DeepMind names a range of uses in its post, and most of them have one thing in common: they replace a world that is expensive or slow to build by hand.
- Robots and agents. DeepMind talks about "developing training environments for robots and agents." The model page mentions SIMA, its game-playing agent, and Gemini Robotics. Training in a generated world is cheaper than training in a warehouse.
- Film. DeepMind says filmmakers can run several takes of a scene in a live simulation. That is previsualisation, the rough version you make before you spend money on the real one.
- Games. Developers can prototype an interactive idea from a written description and play it the same afternoon.
- Testing. DeepMind says its testers come from "creative industries, journalism, education, gaming, and engineering." Engineering teams can use generated worlds to see how an agent behaves in situations nobody scripted.
Every item on that list is a rehearsal for something else. None of them is the thing a customer pays for every day.
What can Genie 4 not do yet?
DeepMind is honest about the limits, which I appreciate. Its post says "visual details and object permanence can still shift or drift over time," especially in longer sessions or after several edits. The model page adds that it cannot yet simulate real places with full accuracy, that legible text usually appears only when you ask for it, and that real-time audio accuracy is still being improved.
Combine that with "a few minutes of continuous interaction" and you have a clear picture. Genie 4 is good for short, rich, surprising scenes. It is not yet a place where state stays put. A door you opened may not stay open. A number on a sign may change.
On safety, DeepMind says all video and audio from Genie 4 carry SynthID, its watermark for AI-generated content. That is useful for anyone who publishes what they generate.
Should you rebuild your app around a world model?
No. And I would push back on anyone who suggests it this year.
A world model is a simulator. It is very good at producing a believable next moment. A customer application is a different thing. It has to remember an order, apply a discount correctly, keep a medical record intact and pass an audit. Those jobs need state that never drifts. That is exactly what Genie 4 says it cannot promise.
When a new model looks like magic, there is pressure to put it at the centre of the product. Usually the right move is smaller. Put it next to the product, where its strengths help and its weaknesses cost nothing. A simulator that drifts is fine for rehearsing. It is a problem in a checkout.
Where would I put it instead?
- In the test suite. If you build agents, generated worlds give you unusual cases to test against, cheaply.
- In discovery. If your team designs spatial or interactive products, a few minutes in a generated scene can replace a week of mock-ups for an early conversation.
- In training data. If you train robots or vision models, simulation is where generated worlds already earn their keep.
What would I do this week?
Join the waitlist if your work involves agents, robotics or interactive media. Pick one narrow question you want answered, such as "can our agent cope with a scene it has never seen," and test only that. Keep the experiment away from your production roadmap until there is a public release, a price and a clear statement on how long a world stays consistent.
Where Alongside fits
Deciding whether a world model belongs in a product, or stays a research experiment, is the kind of work covered by Alongside's product development.
