Type a Sentence. Walk Around Inside It. Then Change the Weather.
That’s the pitch behind Google Genie 4, announced October 27 by Google DeepMind — and unlike most AI announcements this year, the pitch is almost literally what the thing does. You type a sentence, or drop in an image, and Genie 4 generates a world you can move through in real time: walk down a street, turn a corner, watch people in the scene move and react. Then, without stopping or restarting, you type another sentence and change the weather while you’re standing in it.
A video generator makes a clip you watch. A world model makes a place you inhabit. That distinction is the whole story here, and it’s why Genie 4 matters even though almost nobody can use it yet.
Three things are actually new
First, sound — and not as an afterthought. Genie 4 is DeepMind’s first real-time world model with native audio, meaning the music, ambient noise, and spoken dialogue are generated simultaneously with the world itself, not dubbed on afterwards. Characters in a scene can converse and react in real time, to you and to events around them. A world you can talk to is a different medium from a world you can only watch, and it’s the single biggest change from Genie 3.
Second, duration. Coherent scenes now last up to about two minutes, double Genie 3’s 60-second limit. Two minutes still sounds short, and it is — but it’s long enough for a scene to develop. A character can cross a room. Weather can turn. A conversation can go somewhere. The reason world models have always been capped so short is drift: the longer a simulation runs, the more small details slip. DeepMind is candid that this limit hasn’t gone away. In its own words, visual details and object permanence can shift or drift over time, especially in longer runs, after multiple complex prompt edits, or while navigating. That honesty matters, because it’s the gap between a demo and a tool.
Third, steering. Mid-simulation, in plain language, you can change the weather, move the camera, add a person or an object, and the scene keeps running. You navigate with keyboard controls, and a guiding system adapts your prompts to fit what’s already happening. Under the hood, two pieces of the Gemini family do the heavy lifting: Gemini Omni anchors the simulations in real-world appearance and behavior, so objects look and act like themselves, while Gemini 3.8 Flash acts as the in-loop guiding agent, breaking down your instructions and checking each frame to keep the simulation coherent. DeepMind also points to stronger simulated physics than before — better collisions and aerodynamics, for example — which is what makes a generated world feel like a place with rules rather than a screensaver.
Who gets it, and who doesn’t (yet)
For now, Genie 4 is a research preview for trusted testers — DeepMind names filmmakers, game developers, and robotics teams as the first group. Wider access is planned in stages: Google AI Ultra subscribers through Project Genie, and early-access partners through the Gemini API and Google AI Studio. A waitlist is open for everyone else. DeepMind has given no public launch date or API price, so anyone quoting you a date for general availability is guessing.
That staged rollout is also a clue about who this is for first. The use cases DeepMind keeps returning to are film previsualization — blocking out a scene before anyone builds a set — rapid game prototyping, and training environments for robots and agents, where an infinitely variable simulation beats a hand-built one. Those are professional workflows, not consumer toys, and the current evidence is mostly company demonstrations and early tester feedback. A convincing generated scene does not by itself prove dependable physics or repeatable behavior, and DeepMind’s own drift warning is the reason to keep that distinction in mind.
It’s also worth placing Genie 4 in its field. DeepMind is not alone here: NVIDIA’s Cosmos, Runway’s GWM Worlds 2, and ByteDance’s reported Seedance are all chasing the same idea of simulated, navigable worlds. What sets Genie 4 apart on paper is the combination — photorealism, native audio, and live steering in one real-time system — rather than any single feature. And there’s a provenance detail worth noting: DeepMind says Genie 4 outputs will carry an invisible SynthID watermark, though it hasn’t explained how detectable that mark remains after editing or re-encoding.
The honest version of what happens next
Genie 4 is best understood as a direction, not a destination. The demos are genuinely striking — worlds that look photographic, sound alive, and respond when you poke them — but the product, for most people, is a waitlist. What to watch is concrete: when Project Genie opens to Ultra subscribers, what the first independent testers report about drift and control, and whether any game or film team ships something real built on it rather than just a demo reel.
If those things land, the implications are bigger than a neat demo. A world you can generate, steer, and talk to collapses the distance between imagining a scene and standing inside it — for prototyping a game level in an afternoon, for rehearsing a shot before a shoot, for training a robot in a thousand kitchens that don’t exist. If they don’t land, Genie 4 joins a long line of impressive world models that stayed in the lab. Either way, the two-minute, talking, steerable world is now the bar every rival has to clear, and Google set it in public this week.
Related: Meta’s Muse Spark Open — another open frontier in AI models. GPT-6.1 Terra — OpenAI’s computer-use model from the same week. Claude Duet — Anthropic’s proactive agent. And Apple’s own big launch this week: the touchscreen MacBook Pro.
Leave a Reply