Google DeepMind announced Genie 4 on October 27, 2026, and it is a different kind of AI from the chatbots and video generators most people know. Instead of answering a question or rendering a finished clip, Genie 4 builds a world you can step into: a place you explore with your keyboard, where characters move and talk, sounds happen on their own, and you can change the scene mid-stream just by typing what you want.
It is still a research preview, open only to a small group of trusted testers. But Genie 4 matters because it shows where AI is heading next — from generating content to simulating reality in real time. Here is what it actually does, how it works, who can use it, and where it still falls short.
What Is a World Model, Anyway?
Most AI tools you have used so far produce a finished thing: a paragraph, a picture, a video. A world model works differently. It simulates a place frame by frame, predicting what should happen next based on what you do. Walk toward a door and the room beyond appears. Knock over a barrel and it rolls, collides, and settles. Type “make it rain” and the sky darkens while the characters react.
That is the idea behind Genie 4. You start with a text prompt or an image, and the model generates an explorable, photorealistic environment in real time. You move through it with standard keyboard controls and a camera, the way you would in a video game — except nothing here was built by a game studio. The whole world is being imagined, moment by moment, by the model.
If that sounds familiar, it is because Genie 4 is the fourth step in a line of research. Genie 3, released as a research preview in August 2025, could hold a scene together for about a minute at 720p and 24 frames per second. Genie 4 roughly doubles the coherent run time to about two minutes, and adds two things its predecessor lacked entirely: sound and live steering.
The Two Big Upgrades: Sound and Live Steering
Native audio, generated in the same loop
Genie 4 is DeepMind’s first real-time world model that generates audio natively. That means the sound is not pasted on afterward. Wind in the trees, waves on a beach, background music, and spoken dialogue are all produced inside the same simulation loop as the visuals.
The dialogue is the part that changes the feel of the whole thing. Characters in a Genie 4 world do not just stand around waiting for you — they talk back, reacting to what you do and to events in the world as they happen. A beach scene arrives with crashing waves; a character you approach can answer you in real time. For anyone who has tried an AI video tool, the difference is obvious: those tools hand you a silent clip, while Genie 4 hands you a place that sounds alive.
This is worth keeping in perspective alongside the other AI tools people actually use today. Text-to-video generators, like the ones covered in our roundup of the best free AI video generators in 2026, produce finished clips you watch. Genie 4 produces a world you inhabit. The two are related, but they solve different problems.
You can reshape the world while it is running
The second upgrade is steering. In earlier world models, once a scene started, you were largely stuck with it. In Genie 4, you can type a change while the simulation is still running: change the weather, move the camera, swap an object, or drop a new person into the scene. The model keeps going, adapting your prompt to fit what is already on screen.
This matters more than it sounds. It turns world generation from a one-shot gamble — generate, watch, discard, repeat — into something closer to directing. A filmmaker can test a lighting change without re-rendering a whole scene. A game designer can try a different layout without rebuilding a level. A robotics team can throw an unexpected obstacle into a training scenario and watch how an agent copes.
How Genie 4 Works Under the Hood
Genie 4 is not one model working alone. It is a small team of models, each handling a different job:
- Gemini Omni anchors the world in how things really look and behave. It is the video-model side of the system, responsible for the appearance of people and places and for keeping movement believable.
- Gemini 3.8 Flash acts as the guiding agent inside the loop. It watches the scene, breaks a goal into a sequence of steps, checks each new frame to confirm the step actually happened, and replans when you intervene or the world shifts.
- Gemini 4 Argon (paired in DeepMind’s demonstrations) breaks complex, multi-step tasks into steps that Genie 4 then simulates in real time. One demo showed it splitting a game of checkers into steps and rendering a humanoid robot carrying them out.
The physics have improved as well. DeepMind cites better aerodynamics and collisions, and calls the simulated people the biggest leap: characters that move naturally, respond to their environment, and can be steered interactively. Whether that holds up outside of DeepMind’s own curated demos is something independent testers will have to confirm — so far, every claim about Genie 4 comes from Google itself.
This layered design is part of a broader shift in how AI systems are built. Instead of one giant model doing everything, several specialized models cooperate, each checking the others’ work. It is the same direction visible in OpenAI’s recent work, where GPT-6’s Intelligent UI turns chat replies into interactive tools rather than plain text. The interface is becoming the application.
What Genie 4 Is Actually For
DeepMind frames world simulation as a crucial step toward more capable AI, and the practical uses fall into a few clear buckets:
Training robots and AI agents safely. A robot that learns in a simulated warehouse can make a thousand mistakes without breaking anything. Genie 4’s interactive worlds give agents a place to practice complex, multi-step tasks — navigating, manipulating objects, responding to people — before they ever touch real hardware. DeepMind explicitly names this as a use case, and it is probably the most commercially important one.
Filmmaking and pre-visualization. A director can block out a scene, test camera angles, and change weather or lighting in real time instead of waiting on renders or reshoots. The two-minute coherence limit means this is a sketching tool, not a production pipeline — but for storyboarding and pitching, a living sketch beats a static one.
Game prototyping. Level designers can generate a playable space from a sentence, walk through it, and reshape it on the fly. The gap between “idea” and “something you can play” collapses from weeks to minutes.
Education and journalism. Imagine walking through a reconstruction of a historical site, or a simulated environment that shows how a scientific process unfolds. That is further off, but it is the kind of use DeepMind points to when it talks about amplifying creative tools.
All of this fits a pattern that is already reshaping daily life. As we covered in our look at how AI is changing everyday life in 2026, AI is moving from something you consult to something you inhabit — assistants that act, interfaces that adapt, and now worlds that respond.
Who Can Use Genie 4 Right Now — and How to Get Access
Honestly? Almost nobody, yet. Here is the rollout as DeepMind has described it:
- Trusted testers first. The research preview is currently limited to a small group of filmmakers, game developers, and robotics teams.
- Google AI Ultra subscribers next. Genie 4 is coming to Project Genie, Google’s web app for building worlds, for AI Ultra subscribers (the $249.99-per-month tier) in the coming weeks.
- Developers after that. Early-access partners will get it through the Gemini API and Google AI Studio.
- Everyone else: a waitlist. Google has not announced a general release date or a standalone price.
If you want to be ready when access opens up, the practical steps are simple: join the waitlist through Google’s Project Genie pages, and if you already subscribe to Google AI Ultra, watch for the Project Genie rollout. Do not pay a third party claiming to sell Genie 4 access — there is no official reseller channel, and anything offered that way is a scam.
The Limits DeepMind Admits — and the Ones It Does Not
DeepMind is unusually candid about Genie 4’s weaknesses, and they are worth taking seriously before you get carried away:
- Two minutes of coherence, then drift. Visual details and object permanence can shift as a session runs longer, after several complex prompt edits, or while you navigate. Objects can vanish; details can morph.
- Readable text is unreliable. Signs, labels, and documents inside generated worlds often come out garbled — the same weakness text-to-image tools have had for years.
- Real-time audio accuracy is unresolved. Dialogue and sound effects are generated live, and they do not always match the scene perfectly.
- Complex physics and spatial accuracy remain research problems. Collisions and aerodynamics are better, not solved.
There are also gaps DeepMind does not highlight. Unlike Genie 3, Google has published no resolution, frame rate, latency, or benchmark scores for Genie 4 — only demo clips and the “up to two minutes” coherence figure reported by TechCrunch. And one headline claim deserves a footnote: Google calls Genie 4 the first real-time world model with native audio, but Runway’s GWM Worlds 2, released a couple of months earlier, is also playable and also generates audio. “First” here is Google’s claim, not an established fact.
There is one more safeguard worth noting: every video and audio clip Genie 4 produces carries SynthID, Google’s imperceptible watermark marking it as AI-generated. That matters as these worlds get more photorealistic, because the line between a simulated scene and a real recording is getting harder to see with the naked eye.
Genie 4 vs. the AI Tools You Already Use
It helps to place Genie 4 next to tools people already know:
| Tool | What it makes | Can you steer it live? | Sound? |
|---|---|---|---|
| Text-to-video generators | A finished clip | No — regenerate from scratch | Usually added after |
| Chatbots like ChatGPT or Gemini | Text, images, tools | Conversationally, yes | No |
| Genie 4 | A world you walk through | Yes — type changes mid-scene | Yes, generated live |
The comparison that matters most is with plain prompting. Writing a good prompt is still the core skill — it is the same craft behind the AI image generation techniques that turn text into pictures. Genie 4 just moves that skill into a live environment, where a prompt is not a one-time instruction but an ongoing conversation with a world.
And for Gemini users specifically, Genie 4 sits alongside a fast-moving product line. Google is also retiring Gemini Gems in favor of Skills, as explained in our guide to Google retiring Gemini Gems — a reminder that Google’s AI features change quickly, and today’s preview may look very different by general release.
Should You Care About Genie 4 Yet?
If you are a filmmaker, game developer, or robotics researcher, yes — join the waitlist, because this is a tool that could change how you prototype. If you are a curious tech user, the honest answer is: watch it, don’t chase it. The preview is closed, the two-minute limit is real, and the most impressive demos are Google’s own.
But the direction is unmistakable. AI started by answering questions, then by making pictures and videos. Genie 4 points at the next step: AI that builds places — interactive, audible, steerable places — and lets you walk around inside them. The first version is a two-minute sketch. The versions after it will be longer, sharper, and open to everyone. When they arrive, the people who understood the idea early will be the ones who know what to build with it.
Frequently Asked Questions
Is Google Genie 4 available to the public?
Not yet. As of October 30, 2026, Genie 4 is a research preview for trusted testers only. Google says it will reach Google AI Ultra subscribers through Project Genie and early-access developers through the Gemini API and AI Studio in the coming weeks. No general release date has been announced.
How long can a Genie 4 world last?
About two minutes of coherent simulation, roughly double Genie 3’s one-minute limit. After that, visual details and object permanence can start to drift.
Does Genie 4 generate sound?
Yes — it is DeepMind’s first real-time world model with native audio, including ambient sound, music, and interactive dialogue generated alongside the visuals.
How much will Genie 4 cost?
Google has not announced pricing. The first public route is expected to be through the Google AI Ultra subscription, which currently costs $249.99 per month, but Genie 4 access within that tier has not been separately priced.
Is Genie 4 a game engine?
No. It generates playable worlds from prompts, but it is a research model with a two-minute coherence limit, not a production game engine. Its most immediate value is prototyping and agent training, not shipping finished games.
