What Is Google Genie 4? The World Model That Talks Back
Google Genie 4 is a world model: an AI that does not answer questions but builds a place you can walk through — and now, for the first time, a place you can change while you are inside it. On November 18, 2026, Google DeepMind opened Genie 4 to eligible Google AI Ultra subscribers inside a redesigned Project Genie experience. Type a prompt mid-session and the weather shifts, a character appears, the story turns — without the world restarting. And the world talks back: Genie 4 generates real-time audio alongside its visuals, from ambient sound to character dialogue that reacts to what happens. This guide explains what is Google Genie 4, how it differs from a chatbot and from its predecessor Genie 3, what changed on November 18, and what it can — and cannot — do yet.
What is a world model?
A world model is an AI system that simulates an environment rather than producing an answer. Give a chatbot a prompt and it returns text; give a world model a prompt and it returns a navigable scene — a street, a forest, a manor house — rendered frame by frame as you move through it, with objects, physics and events that respond to what you do. The model has to keep track of where things are, how they behave, and what has already happened, and generate the next moment in real time.
That makes world models a different kind of AI from the language models most people know. A language model predicts the next word; a world model predicts the next moment of a place. Google DeepMind has been building this line of research since the original Genie in 2024, through Genie 2 and Genie 3, and Genie 4 is the latest and most realistic version. The practical payoff Google points to goes beyond games: simulated worlds can serve as training grounds where robots and AI agents learn to navigate, plan and react before they are deployed in the real world. The entertainment side — explorable worlds you can build, share and remix — is the part the public can now try.
What Genie 4 does
Genie 4’s headline capability is editing a world in the middle of a session. Earlier world models worked like this: you described a world, the model generated it, and if you wanted something different you started over. Genie 4 lets you type a new prompt while you are already navigating — change the weather, drop in a character, redirect the action — and the world adjusts around you. Google describes this as the model responding to prompts mid-session, and it is the single biggest change from Genie 3.
The second headline is sound. For the first time in this model line, visuals come paired with real-time native audio that adapts as the scene unfolds. That means ambient soundscapes that match the environment, and character dialogue that reacts to the user and to events in the simulation, generated together with the picture rather than dubbed on afterwards. Google calls Genie 4 its most realistic and interactive world model yet, citing higher-fidelity visuals and physics alongside the audio.
Under the surface, Genie 4 is natively multimodal. According to DeepMind’s research description, Gemini Omni provides a real-world reference for how simulations should look and behave, while other Gemini models supply the reasoning that keeps a world coherent over time. One of them, Gemini 3.8 Flash, can act as an in-loop guiding agent: it watches the scene, breaks goals into steps, checks incoming frames for progress, and changes its plan when the world or the user’s input shifts. In other words, part of what makes Genie 4 feel responsive is another AI model steering it from inside.

What changed on November 18 — and what happened on October 27
Genie 4 itself is not brand new. Google DeepMind first announced it on October 27, 2026 as a research preview for trusted testers, with Ultra subscribers in Project Genie named as the next group to get access. November 18 is when that promise landed: first-time public access to Genie 4, inside a redesigned Project Genie experience, for eligible Google AI Ultra subscribers. Pro members are not in the first wave — they can sign up for the waitlist on that same page instead.
The November 18 update is bigger than one model. It redesigns Project Genie itself — the browser-based prototype Google first rolled out on January 29, 2026, originally powered by Genie 3 — around four things: Genie 4, upgraded Genie 3, a creation agent, and a community gallery. So the news is not only that Genie 4 is available, but that the whole experience around it has been rebuilt to make worlds easier to build and share.
Genie 3 vs Genie 4
| Genie 3 | Genie 4 | |
|---|---|---|
| Visual and physics fidelity | High | Highest — Google’s most realistic world model yet |
| Mid-session prompting | Real-time in-world events added Nov 18 | Full mid-session editing: weather, characters, action |
| Native audio | No | Yes — real-time ambient sound and character dialogue |
| Session length per generation | Up to 3 minutes (up from 60 seconds) | Shorter runs; fidelity over duration |
| World chaining | Yes — link multiple worlds into longer journeys | Not the focus; Genie 3 handles longer journeys |
| 360-degree panorama seeding | Yes — seed worlds from panorama images | No |
| Best for | Longer runs, panorama-seeded spaces | Highest-fidelity, most interactive worlds |
The two models are not a simple old-versus-new replacement; Google keeps both in Project Genie’s model picker because they do different jobs. Genie 4 is the choice when fidelity matters most — the sharpest visuals, the most responsive physics, and the only one with native audio. Genie 3 is the choice when duration matters: its sessions now run up to three minutes per generation, triple the old 60-second limit, and worlds can be chained together for longer journeys. It also remains the only one that can seed a world from a 360-degree panorama image. The sensible way to think about it is that Genie 4 is the showpiece and Genie 3 is the workhorse — and a user can switch between them in the same experience.
The Project Genie agent: describe it, don’t build it
Before this update, making a world meant writing a text prompt or uploading an image and accepting what came back. The new Project Genie agent turns creation into a conversation. Describe a vision in natural language and the agent orchestrates Google’s multimodal models to craft the world’s opening visuals — establishing its signature look — paired with a custom background score tailored to the mood you described. You do not assemble the world piece by piece; you describe it, and the agent does the assembly.
Gemini models stay involved after creation too. While you explore, a Gemini model can observe what is happening on screen, surface context-aware suggestion chips for where to go next, or trigger in-world events automatically as the journey unfolds. This is where Genie 4 connects to the wider Gemini family: the world model renders the place, and a language model reasons about what should happen in it. Readers who want the background on that reasoning layer can see our guide to agentic AI, which explains how AI agents plan and act. And Genie 4 arrives alongside other November releases from Google — including Gemini 4 Hydrogen and Gemini 4 Neon — as the company pushes its newest models into products at the same time.
The gallery: worlds to explore, remix and publish
The new community gallery is the social layer. It mixes curated worlds from Google with creations from the community, and every world in it can be explored and remixed. The launch examples show the range Google has in mind:
- Living Memory — turn personal photos into explorable spaces, so a picture becomes a place you can walk through.
- Dino Field Guide — wander prehistoric habitats grounded in 360-degree panoramas.
- Rock Paper Scissors — pit themed objects against each other on a desk, with Gemini acting as real-time referee.
- Villa Giallo — an atmospheric, room-by-room mystery inside a gothic manor, built with creative technologist Will, also known as BLVCKL!GHT.
- The Life of Ants — navigate a complex colony from an ant’s perspective, built with filmmaker Anand Gandhi and drawing on Genie 4’s photorealism.
Remixing works through the agent: take an existing gallery world and tweak its structure, soundtrack or rules to make it your own, then publish your version back to the gallery for others to explore and remix in turn. That publish-and-remix loop is what turns Project Genie from a demo into a platform — the value grows as more people build on each other’s worlds.

How to try Genie 4
Access starts with a Google AI Ultra subscription: eligible Ultra subscribers can try the new Project Genie experience from today, in the browser, at the Project Genie page linked from Google’s announcement. Pro members are not in the first wave — they can sign up for the waitlist on that same page instead. There is no separate price published for Project Genie access — it comes with a Google AI Ultra subscription, and no dollar figure for that subscription is stated here because none was verified from an official Google page.
A few practical points before you open it. The original Project Genie prototype was limited to users in the United States aged 18 and over; Google’s November 18 announcement describes the rollout to “eligible” Ultra subscribers without restating a country list, so availability may vary by region and you should check the Project Genie page for your account. Sessions are short — Genie 3 runs up to three minutes per generation, and Genie 4’s runs are shorter still — so think of a visit as a tour of a world, not an afternoon inside one. And because everything is generated live, results vary: the same prompt can produce a different world on a different day.
Limits and open questions
- Worlds drift. Visual details and object permanence can change during longer sessions, after complex edits, or while navigating. A door you saw a minute ago may not be where you left it. Google names long-term scene permanence as an unresolved challenge.
- Camera latency. There is a delay between what you do and what the world shows, which Google lists among the key challenges it is still working on. Fast, precise movement is not yet reliable.
- Physical interactions are limited. Intricate interactions between objects — picking things up, manipulating them precisely — remain an open problem, per Google.
- Audio is not yet precise. Real-time audio precision is on Google’s own list of unresolved challenges, so dialogue and sound may not always match the scene cleanly.
- Text and spatial accuracy. Legible text inside generated worlds and accurate spatial relationships are still active research areas, according to independent reporting on the launch.
- It is a research prototype. Google calls Genie 3 and Genie 4 frontier research models, not finished products. Expect rough edges, changing behaviour, and features that appear and disappear as the research moves.
- Watermarking is built in. All video and audio generated in Project Genie carries SynthID watermarking, and all inputs and outputs pass through safety screenings. Anything you make is marked as AI-generated by design.
Who is Genie 4 for?
Consider trying it if you are curious about where AI is heading beyond chatbots, you make games, films, or interactive art and want a fast way to sketch a world, or you are a researcher or developer interested in simulated environments for training agents and robots. The gallery alone is worth a visit: it is the fastest way to understand what a world model is, because you experience one instead of reading about it.
Wait if you need a stable tool for finished work. Worlds drift, sessions are short, and the model is a research prototype. A game developer prototyping a level will get inspiration from Genie 4, not a shippable asset. And if you are on Google AI Pro rather than Ultra, the waitlist is your only route for now.
Look elsewhere if what you actually want is a chatbot, a writing assistant, or an image generator. Genie 4 does none of those jobs; its Gemini siblings do. The point of a world model is the place, not the answer.
Frequently asked questions
What is Google Genie 4?
Genie 4 is Google DeepMind’s most realistic and interactive world model. It generates navigable, simulated worlds you can explore in a browser, edit mid-session with text prompts, and hear through real-time native audio. It opened to eligible Google AI Ultra subscribers in Project Genie on November 18, 2026.
How is Genie 4 different from a chatbot?
A chatbot returns text answers. Genie 4 returns a place: a simulated environment rendered frame by frame as you move through it, with objects, physics and events that respond to you. It is closer to a game engine driven by AI than to a conversation.
What is the difference between Genie 4 and Genie 3?
Genie 4 has higher-fidelity visuals and physics, full mid-session editing, and real-time native audio — but shorter sessions. Genie 3 runs up to three minutes per generation, supports world chaining and 360-degree panorama seeding, and remains the better choice for longer journeys.
How do I access Genie 4?
Through Project Genie, in the browser, with an eligible Google AI Ultra subscription. Google AI Pro members can join a waitlist on the same page. No separate price for Project Genie access has been published.
Can I change a world while I am inside it?
Yes — that is Genie 4’s headline feature. Type a new prompt mid-session to shift the weather, drop in a character or redirect the action, without restarting. Genie 3 can also trigger real-time in-world events with natural-language prompts.
Is Genie 4 a finished product?
No. Google calls Genie 3 and Genie 4 frontier research models. Worlds can drift over long sessions, camera latency and precise physical interactions are unresolved, and text and spatial accuracy are still being worked on. All generated video and audio carries SynthID watermarking.