Google's Genie 4 Worlds Now Talk Back, and You Can Rewrite Them While You Walk Through Them
A video you watch is finished. A world model is not.
On Tuesday, October 27, Google DeepMind released Genie 4, the newest version of its world model, as a research preview for trusted testers. Type a sentence or drop in an image, and Genie 4 builds a place you can move through in real time, using keyboard controls, frame by frame, as you go. That is the difference between a world model and a video generator. A video model hands you a clip and you watch it. A world model hands you a scene and answers when you act: turn left, and it has to work out what is around the corner, right now, with no time to render anything in advance.
Genie 4 adds three things that move this from a clever demo toward a usable tool. It generates sound natively, including music, ambient noise and characters who talk and react. It lets you rewrite the world while it is running, by typing a plain-language prompt mid-simulation. And it holds a scene together for up to about two minutes, double the roughly one minute reported for Genie 3, which DeepMind released as a research preview a little over a year ago. None of that makes it a game engine or a finished product, and DeepMind says so itself. But the direction is clear, and the competition around it is starting to consolidate fast.
What is actually new
- Native audio, generated with the world, not added after. Genie 4 is DeepMind’s first real-time world model to produce sound as part of the simulation itself: wind and waves, a music bed, and spoken dialogue from characters in the scene, generated at the same time as the visuals rather than in a separate pass. Characters can converse with each other and react to you and to events in the world, in real time.
- Live steering by prompt. While a simulation runs, you can type instructions such as changing the weather, moving the camera, or bringing in a new person or object, and the world adjusts without stopping. Sundar Pichai demonstrated the idea with a motorcycle on a coast road: prompt that the traffic light turns green, and the scene works out what follows, the engine revs and the bike pulls away.
- Longer coherence. Scenes stay consistent for up to about two minutes, double Genie 3’s limit, with better physics and more humanlike characters, according to TechCrunch’s briefing. Two minutes is still short. It is long enough, though, for something to happen: a character can cross a space, weather can turn, a short conversation can go somewhere.
- Better people and physics. DeepMind says its biggest leap is in human characters, who move more naturally and can be steered in real time, and in physical behaviour such as collisions and how objects respond when you interact with them.
The quiet change is the director inside the loop
The most important detail in the announcement is not the audio. It is that Genie 4 does not run alone. DeepMind has wired its Gemini models into the simulation loop itself. Gemini Omni anchors the world in real-world appearance and behaviour, grounding how things look and how they should act. Gemini 3.8 Flash runs alongside as a guiding agent: it watches the scene as Genie generates it, breaks what should happen next into steps, plans, and checks that the steps actually happened, adjusting as the world changes between your inputs.
That pattern matters beyond pretty scenes. A world model with a reasoning model inside it is a simulator that can be directed, not just prompted. DeepMind showed one use directly: in a research experiment tied to Gemini Robotics, a Gemini 4 Argon model broke a robot task down into steps, and Genie 4 simulated those steps in real time. If that pairing works reliably, a robot can practise a task in a generated world that responds and sounds right, fail safely a thousand times, and only then try it in the real one. Training environments for robots and agents, not entertainment, may be the first place this earns money.
Who can use it, and on what terms
Almost nobody, yet. The honest access list is short:
- Now: trusted testers only, in a research preview.
- Coming soon: Google AI Ultra subscribers, through Project Genie, and early-access partners in the Gemini API and Google AI Studio. A waitlist is open for Project Genie.
- Not published: no price for Genie 4 itself, no release date for general access, and no public figures for resolution, frame rate, inference cost or maximum session length.
Everything Genie 4 produces, video and audio, carries a SynthID watermark, DeepMind’s imperceptible marker for AI-generated content. In a medium built to look real and respond like the real world, that provenance layer is not a footnote. It is part of the product.
The limits DeepMind lists itself
Credit where it is due: the announcement is candid about what still breaks. Visual details and object permanence can shift or drift over time, especially in longer runs, after multiple complex prompt edits, or while navigating around a scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are all listed as active research problems, not solved ones.
Read that plainly. A world that forgets what an object looked like, or moves it, is a sketchpad that moves, not a replacement for a game engine, a film pipeline or a physics simulator. Two minutes of coherence is a real doubling and a real constraint at the same time. Anyone planning a product on top of world models today should plan around drift, short sessions and hosted-only access, because those are the terms on offer.
Why the timing matters: the field is consolidating
Genie 4 does not land in an empty field. TechCrunch notes that Runway released its second world model, GWM Worlds 2, a couple of months ago, a playable model that responds to user inputs and also generates audio. At the same time the industry is consolidating: Fei-Fei Li’s World Labs is being acquired by AMD, Amazon is in talks to buy Decart, and Rivian has acquired Third Dimension to bring simulation in-house for self-driving testing.
That pattern tells you what world models are actually for, in the eyes of the people spending money. The buyers are a chip company, a cloud giant and a car maker. Two of the three are buying simulation for machines that have to act in the physical world. Games and film prototyping are the visible use case, and trusted testers are already pointing to storyboarding and pre-visualisation. The larger prize is a cheap, endless, safe place to train and test robots, vehicles and agents before they touch anything real.
What to do with this, this week
- If you build games, film or interactive content: join the Project Genie waitlist and treat this as a prototyping tool. Two coherent minutes is enough to test a scene, a mood or a mechanic. It is not enough to ship a world, and any plan that assumes public API pricing exists today is speculation.
- If you work in robotics or simulation: watch the Gemini-plus-Genie pairing, not the demos. A reasoning model directing a world model that simulates its plan is a training loop. Ask any vendor pitching it for drift rates, session limits and cost per simulated hour, the numbers nobody published this week.
- If you buy AI tools generally: note the release shape, because it is becoming the standard one. Research preview first, paid-tier subscribers next, API partners after that, watermarking on by default, and the hardest limits disclosed in the fine print rather than hidden. Budget and timelines should follow that shape, not the headline.
The bottom line
Genie 4 makes generated worlds hearable, talkable and steerable while they run, and it puts a reasoning model inside the simulation to keep it on track. Those are real advances, delivered inside real limits: about two minutes of coherence, drift that DeepMind names openly, and access that almost nobody has yet. The race now is not who can generate the prettiest world. It is who can keep a world consistent, physical and cheap for long enough that a robot, an agent or a paying customer can actually use it. Google just doubled its time on that clock.