Google has opened the door to its most realistic world model yet. On 18 November 2026, Google DeepMind rolled Genie 4 into Project Genie, its browser-based world builder, for eligible Google AI Ultra subscribers — and the headline feature is not the world itself, but the fact that you can change it while you are standing inside it.
Type a prompt mid-session and the weather shifts, a character walks in, or the action turns a different way, without starting over. For the first time in the series, the visuals come with real-time native audio that adapts as the scene unfolds, from ambient soundscapes to character dialogue. That combination — live editing plus sound — is what separates Genie 4 from every world model Google has shipped before it.
What Genie 4 actually is
Genie 4 is DeepMind’s most advanced interactive world model: a system that generates a 3D space an AI — or a person — can move through, rather than a video you watch. Google describes it as its “most realistic and interactive world model yet”, and the claim rests on three things: higher-fidelity visuals and physics than Genie 3, the ability to steer the world with typed prompts while you explore it, and sound generated alongside the picture.
Under the surface, Genie 4 is built on the Gemini family. Gemini Omni anchors the simulation in how the real world looks and behaves, while other Gemini models supply the reasoning that keeps a scene coherent from one moment to the next. In Project Genie, Gemini 3.8 Flash can even act as an in-loop guiding agent: it watches the scene, breaks goals into steps, checks frames for progress, and changes its plan when the world — or your prompt — shifts. That is the part aimed at researchers: Google presents these worlds as training environments for robots and agents, not just as a creative toy.
The model itself is not new today. DeepMind announced Genie 4 on 27 October 2026 as a research preview for a limited group of trusted testers. What changed on 18 November is access: paying subscribers can now try it inside Project Genie for the first time.
What changed in Project Genie on 18 November
The 18 November update, announced by Google DeepMind’s Diego Rivas and Google Labs’ Pavan Kumar, is a rebuild of the whole Project Genie experience, not just a model swap. Five pieces arrived together:
Genie 4 in the model picker. The headline addition: prompt mid-session to edit your world in real time, with native audio generated as you go.
Genie 3 gets longer and more flexible. Genie 3 stays available for longer runs and panorama-seeded spaces. Sessions now run up to 3 minutes per generation, up from 60 seconds, and you can link multiple worlds together into longer journeys. You can also seed worlds from 360-degree panorama images and trigger real-time in-world events with natural-language prompts.
The Project Genie agent. Creation is no longer a single prompt and hope. A conversational agent builds the world with you: describe your vision in natural language and it orchestrates Google’s multimodal models to set the look, then adds a custom background score matched to the mood. While you explore, Gemini models watch the screen, surface suggestion chips for where to go next, or trigger events automatically.
A community gallery. You can now publish worlds for others to explore and remix, or start from someone else’s. The launch gallery includes “Living Memory”, which turns personal photos into explorable spaces; “Dino Field Guide”, prehistoric habitats grounded in 360-degree panoramas; and “Rock Paper Scissors”, where Gemini referees themed objects pitted against each other on a desk. Two creator collaborations lead the showcase: “Villa Giallo”, a gothic-manor mystery built with creative technologist Will, known as BLVCKL!GHT, and “The Life of Ants”, an ant’s-eye colony built with filmmaker Anand Gandhi.
Wider access, still gated. Eligible Google AI Ultra subscribers can try the new experience from today; Google AI Pro members can join a waitlist on the same page.
Genie 4 vs Genie 3: which one to pick
The two models now sit side by side in the picker, and the choice is a trade-off, not an upgrade path. Google’s own framing makes the split plain:
| Genie 4 | Genie 3 | |
|---|---|---|
| Best for | Highest-fidelity visuals, physics and sound | Longer runs, panorama-seeded spaces |
| Live editing | Yes — prompt mid-session to change the world | Yes — real-time in-world events via prompts |
| Native audio | Yes — ambient sound, music and character dialogue generated live | No native audio |
| Session length | About a minute of continuous interaction, per DeepMind’s model page | Up to 3 minutes per generation |
| World chaining | Not stated | Yes — link worlds into longer journeys |
| Panorama seeding | Not stated | Yes — seed from 360-degree images |
Pick Genie 4 when the scene itself is the point: a short, dense, atmospheric world where sound and detail matter more than duration. Pick Genie 3 when you want a longer journey, a world built from a real panorama, or several chained worlds. The honest summary is that Genie 4 buys realism and sound for about a minute; Genie 3 buys three minutes without its own audio.
One caution on the numbers: Google’s 18 November announcement does not publish a Genie 4 session length at all. The “around a minute” figure comes from DeepMind’s model page, as reported by i-scoop, and the October preview said scenes stay consistent for up to two minutes. Treat the minute as the working figure, not a guarantee.
Who can try it — and what it costs
Access is the real gate. Project Genie is available to eligible Google AI Ultra subscribers, and Google is rolling Genie 4 out gradually, so not every subscriber sees it on day one. You must be 18 or older and signed in with a personal Google account. Google AI Pro members can join a waitlist, but they are not part of the initial rollout. There is no Genie 4 API.
The price is not in Google’s 18 November post, so it has to be read from Google’s plans pages as reported this week. Project Genie sits on the top Ultra tier — $199.99 a month in the US, according to Google’s May 2026 plans post as reported by TechTimes — not the $99.99 entry Ultra tier. That top tier was cut from $249.99 in May 2026, when the $99.99 tier was added. Prices vary by country — i-scoop reports €219.99 a month in Belgium — and Google changes them without much warning, so check the live price on Google’s plans page before subscribing. Do not pay for Ultra for Genie alone without confirming the tier includes it in your country.
What you can actually do with a minute-long world
A minute sounds like a demo, and for a single unbroken session it is. But the minute is not the product — the edit loop is. Three first things worth trying, in order:
1. Remix a gallery world before building your own. Open “Villa Giallo” or “Dino Field Guide”, change one thing mid-session — the weather, a character, the time of day — and watch how much of the world holds together. That tells you more about Genie 4 than any fresh prompt, because it tests the model’s memory, not its first impression.
2. Turn a photo into a place. “Living Memory” takes a personal photo and makes it explorable. It is the most concrete use case in the launch: a family photo becomes a space you can walk through, rather than an image you look at.
3. Test the agent, not the model. Describe a world in conversation — mood, look, sound — and let the agent set it up. If you like building by talking, the agent is the feature that survives after the novelty of the first world wears off.
Beyond play, the serious uses are the ones Google names: prototyping interactive experiences, testing how an AI agent navigates a space, and building training environments for robots. For those, the minute matters less than the fact that the world reacts.
What Genie 4 cannot do yet
Google is unusually candid about the limits, and they matter more than the showcase. Both Genie models remain, in Google’s words, “world models in the frontier research space”. The open challenges it names are camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence. In practice, that means visual details and object permanence can drift during longer sessions, after complex edits, or while navigating — and legible text and spatial accuracy remain active research areas.
Two more honest limits: all video and audio generated in Project Genie carries SynthID, Google’s imperceptible watermark, and all inputs and outputs go through safety screenings — so this is a walled garden, not a raw model you can take elsewhere. And with no published Genie 4 session length, no API, and a gradual rollout, the experience you get on day one may differ from the one described in the announcement.
Our take: the minute test
Here is the question that decides whether Genie 4 is worth your money: do you want a world, or do you want to change a world while you are in it? If you only want the first, Genie 3’s three minutes, the gallery, and a dozen cheaper image and video tools will serve you better. Genie 4 earns its keep only if you use the live-edit loop — steering a scene, testing an agent inside it, or remixing a gallery world until it becomes yours.
Our verdict: try the gallery first through whatever access you already have, remix one world, and only then decide whether the top Ultra tier is worth it for Genie alone. The technology is genuinely new — a world you can talk to while you walk through it — but at roughly a minute a session and behind a $199.99-a-month gate, it is a research preview with a price tag, not a tool most people need yet. The gallery is the smart way in; the subscription is the commitment.
Frequently asked questions
Is Genie 4 free to try?
No. Genie 4 is available to eligible Google AI Ultra subscribers through Project Genie, rolling out gradually from 18 November 2026. Google AI Pro members can join a waitlist on the same page, but they are not part of the initial rollout. You must be 18 or older and signed in with a personal Google account.
How long does a Genie 4 world last?
About a minute of continuous interaction, according to DeepMind’s model page. Google’s 18 November announcement does not publish a Genie 4 session length, so treat the minute as the working figure. Genie 3, by contrast, now runs up to 3 minutes per generation and can chain worlds together.
Can I still use Genie 3 in Project Genie?
Yes. Genie 3 remains in the model picker for longer runs and panorama-seeded spaces, and the 18 November update extended it to 3-minute generations, added 360-degree panorama grounding, and added real-time in-world events.
What does the Project Genie agent do?
It builds worlds through conversation. Describe your vision in natural language and the agent orchestrates Google’s multimodal models to create the look and a custom background score. While you explore, Gemini models observe the screen, suggest where to go next, or trigger in-world events automatically.
Can I publish or sell the worlds I make?
You can publish a world to the community gallery for others to explore and remix. Google has said nothing about selling worlds or monetising them, so treat the gallery as sharing and remixing only.
Sources and methodology
Primary source: Google’s own announcement, “Project Genie update featuring Genie 4”, by Diego Rivas (Google DeepMind) and Pavan Kumar (Google Labs), published 18 November 2026 and read in full. Independent reporting: Unite.ai, “Google Brings Genie 4 to Project Genie for Eligible AI Ultra Users”, 19 November 2026; TechTimes, “Google Genie 4 Opens to Ultra Subscribers With Live Prompt Edits and No Published Session Length”, 19 November 2026; i-scoop, “Project Genie brings Genie 4 to Europe, and its remix gallery is what organisations should plan for”, 19 November 2026.
All claims were verified against these sources on 19 November 2026. Prices are taken from reporting of Google’s plans pages, not from the 18 November announcement, and vary by country — check Google’s live plans page before subscribing. This article is desk research: no Google AI Ultra subscription was used and Genie 4 was not tested hands-on. Where Google’s announcement and its model page differ — notably on session length — both are stated and the difference is flagged rather than smoothed over.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


