What Is Google's Genie 4? The New AI World Model, Explained (What's New, Access, and the Waitlist)

CodexFuse Team · 2026-11-06T08:03:38.507Z · 27 min read min read

What Is Google's Genie 4? The New AI World Model, Explained (What's New, Access, and the Waitlist)

If you are wondering what is genie 4, the short answer is that it is Google DeepMind's new AI world model, announced on 27 October 2026, that generates an interactive three-dimensional world in real time and lets you walk through it, talk to the people in it, and change it with plain sentences while it runs. Unlike a video generator, it does not hand you a finished clip. It draws the world frame by frame as you act, adds the sound of the place itself, and keeps the scene going as you explore.

Google Genie 4 is the fourth generation of DeepMind's Genie line, and it is the first one with native audio: wind in the trees, waves on a beach, music, and characters who speak back. Everything below was checked against the sources listed at the end on 6 November 2026, ten days after the announcement. As of 6 November 2026, Genie 4 is a research preview for trusted testers only, so this guide explains what Google has actually said and shown, not a hands-on review, because nobody outside Google's trusted testers has used it yet.

Our agentic AI explainer covers the wider idea of AI systems that plan and act on their own, which is the family Genie 4 belongs to. Genie 4 takes that idea into a world you can step inside: the model plans, the world reacts, and you steer both. The rest of this guide walks through what a world model is, what is new against Genie 3, how long a session lasts, how to join the waitlist, and what it costs.

What Is Genie 4?

Genie 4 is Google DeepMind's real-time interactive world model: you type a text prompt, or start from an image, and it generates a navigable 3D environment frame by frame, in real time, that you steer with keyboard controls and natural-language prompts while it runs. DeepMind's announcement puts it this way: Genie 4 "unlocks unprecedented realism and improves real-time interactive simulation", and it is "natively multimodal, enabling audio-like music, ambient noises, and interactive dialogue to be simultaneously generated with the world".

A world model, in one line, is an AI system that uses its understanding of the world to simulate how an environment evolves and how actions affect it, drawing each new frame as you act rather than playing back a pre-rendered video. You can set an event in motion, and Genie 4 simulates what happens next: how the world responds, how it sounds, and how the people in it move, react, and speak. That last part is the point of the whole release: earlier world models were silent dioramas, and this one talks.

Google frames world models as a step toward artificial general intelligence, because understanding and simulating the world is, in DeepMind's words, "a crucial aspect of building AGI". The practical pitch is narrower: safe, infinitely variable simulations for training robots and AI agents, and new interactive experiences for creators.

Genie 4 sits between a game engine and a video generator, and it is neither: a game engine runs rules a designer wrote, a video generator renders a clip you watch, and Genie 4 invents the world as you explore it. That is the genie 4 AI world model in one paragraph: a prompt goes in, a world you can walk through comes out, and the sound comes with it.

What Is New in Genie 4?

Four things are new in Genie 4, and DeepMind structures its whole announcement around them: real-time steering in plain language, more realistic physics and human movement, native audio, and complex multi-step events coordinated by Gemini models. Each one answers a specific weakness of Genie 3, and together they are why Google calls this a new generation rather than an update.

The steering is the change you would notice first. With Genie 4, you can use natural language to modify a simulation in real time, and Gemini contextually adapts your prompt to match the surrounding scene. Between your inputs, Gemini also monitors and steers the scene to keep the world progressing cohesively, so a change you type does not just swap one backdrop for another: the world keeps behaving like itself afterwards.

The physics and movement work is the change you would feel second. Genie 4 is built on an advanced physical understanding of the world, and it simulates real-world physics scenarios such as aerodynamics and collisions. DeepMind calls the human movement "the most dramatic leap": characters move naturally, respond organically to their environments, and can be interactively steered in real time, instead of gliding through the scene like cutouts.

The audio is the headline, because it is the first time any Genie model has had it. Genie 4 is DeepMind's first real-time world model to generate audio natively, and every world carries the ambient sound of its own setting, from wind in the trees to waves crashing at the beach.

All the sounds in DeepMind's demo videos were produced directly by Genie 4, not dubbed on afterwards. The frontier, in DeepMind's own words, is human speech: Genie 4 introduces interactive dialogue that adapts on the fly, so characters converse and react to your inputs or events in the world, all in real time.

Genie 4 vs Genie 3: What Changed?

Genie 4 is a clear step up from Genie 3 in sound, steering, and reasoning, while the basics of the idea are unchanged: both generate interactive worlds in real time from a prompt, and both are research previews rather than products. Genie 3 was announced on 5 August 2025, which makes the gap between the two announcements 448 days, or about one year and three months. The table below compares them on the facts Google has published.

The biggest visible difference is the sound. Genie 3 arrived in August 2025 with silent, minute-long scenes, and every world it made was mute.

Genie 4 generates the soundtrack with the picture: ambient noise, music, and dialogue that reacts to what happens. That is not a cosmetic addition, because a world that talks back changes what the model has to simulate: it must decide what a character would say, when they would say it, and how the scene should sound while they say it.

The second difference is who is doing the thinking. Genie 3 had promptable world events, so you could already change a scene with text, but Genie 4 puts Gemini models in the loop: Gemini Omni anchors the simulations in real-world appearance and behaviour, while other Gemini models provide the reasoning that helps maintain a simulation's coherence and continuity.

Gemini 3.8 Flash serves as a guiding agent in the loop, watching the frames as they arrive, breaking goals into steps, and checking that each step succeeded. DeepMind calls this direction "hybrid world models": Gemini maps the structure and plans the narrative timeline, while Genie renders the immediate physical reality pixel by pixel.

The third difference is what Google has not published. Genie 3 shipped with concrete numbers, 720p at 24 frames per second, and Genie 4 shipped with none.

Google has published no resolution, frame rate, parameter count, or benchmark for Genie 4, so do not carry Genie 3's 720p and 24 fps over to it. That silence is itself a signal: the model is earlier in its life than the confident spec sheet suggests, and the numbers that do circulate come from press briefings, not from Google's pages.

How Long Does a Genie 4 Session Last?

A Genie 4 session lasts about two minutes, according to press reports from the launch briefing, while Google itself only says "a few minutes". That distinction matters, because the two-minute figure is the number everyone quotes, and it does not appear anywhere on Google's own pages. TechCrunch reported from the briefing that Genie 4 can maintain a consistent scene for up to two minutes, double its predecessor's limit, and The Verge reported the same figure, about two minutes, up from one minute for Genie 3.

The "one minute" side of that comparison needs the same care. Genie 3's own announcement said its worlds stay consistent "for a few minutes" at 720p and 24 frames per second, with visual memory extending as far back as one minute ago.

The hard 60-second cap belonged to the Project Genie app, which told users "you have 60 seconds to explore the generated world", not to the Genie 3 model itself. So the honest version of the comparison is: press reports say about two minutes for Genie 4 against a 60-second app cap for Genie 3 in Project Genie, while Google's own wording for both models is the vaguer "a few minutes".

The arithmetic behind those numbers is simple, and it shows why the length matters more than it sounds. A 60-second session at 24 frames per second is 1,440 frames the model must keep coherent; a 120-second session is 2,880 frames, exactly double.

At Genie 3's 720p, that is 22,118,400 pixels every second the model must keep coherent. Every extra frame is another chance for a detail to drift, a person to change clothes, or a doorway to move, which is why DeepMind lists long-term consistency as an open research problem rather than a solved one.

There is one more caveat worth stating plainly. DeepMind engineer Hang Qi reportedly cautioned at the briefing that reaching the two-minute ceiling does not mean the scene stays fully consistent throughout, and DeepMind's own announcement says visual details and object permanence can still shift or drift over time, especially as a simulation runs longer. A two-minute world that slowly forgets itself is still a research result, not a finished product, and Google presents it that way.

How Does Genie 4 Work?

Genie 4 works as a partnership between two kinds of model: Genie renders the world pixel by pixel in real time, and Gemini models watch, plan, and correct what Genie renders. DeepMind's announcement describes the split directly: Gemini Omni anchors the simulations in real-world appearance and behaviour, while other Gemini models provide the complex reasoning needed to help maintain a simulation's coherence and continuity. That is the "hybrid world model" idea DeepMind says it is working toward.

The guiding agent has a name and a job. Gemini 3.8 Flash serves as a guiding agent in the loop: as Genie 4 generates a world in real time, Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track progress and verify that each step succeeded. As the world shifts or the user intervenes, Gemini adapts its plan in the moment, which is what makes multi-step events possible: a character can be told to do something in stages, and the model can notice when a stage fails and try another way.

You steer it two ways at once. Standard keyboard controls handle navigation and camera control, and text prompts steer the scene as it evolves: change the weather, move the camera, introduce a new person or object, and Gemini contextually adapts your prompt to match the surrounding scene. The recordings DeepMind published are real-time interactions, not edited clips, which is the point: the world is being generated while the user is inside it, and the steering happens live.

Our Claude Haiku 5.5 explainer and our Claude Fable 5.1 explainer cover two other recent AI models in the same explainer family, and the pattern is the same in all three: the headline model is only half the story, and the system around it does the planning. With Genie 4, that system is Gemini, and DeepMind is explicit that the two are being brought closer together on purpose.

What Can You Do With Genie 4?

DeepMind tested Genie 4 with trusted-tester cohorts across creative industries, journalism, education, gaming, and engineering, and it named four concrete uses. All of them are tester-stage experiments, not production workflows, and DeepMind frames the work as learning where the model still needs to grow. The table below lists them with the example DeepMind gives for each.

Filmmaking is the clearest example, because it uses everything Genie 4 adds at once. Creatives and technologists are testing Genie 4 as a rapid "draft mode" for pre-visualisation and ideation: a filmmaker can step into a live simulation of a scene and run multiple takes in real time, adjusting camera angles, lighting, dialogue, and action on the fly. That is a rehearsal room you can rewrite with sentences, and it only works because the audio and the steering are native rather than added later.

Journalism is the most surprising example. Photojournalists are exploring how a single 2D archival photograph can become a grounded simulation of that environment in seconds, and teams are testing how populated, dynamic worlds can help people experience historical spaces or visualise how short-term environmental events, like a flash flood or wildfire, unfold. The simulation is grounded in the photograph, not invented freely, which is the difference between a reconstruction and a fantasy.

Game developers get the fastest loop of all. They are testing Genie 4 to go from an initial idea to an explorable prototype in seconds, testing visual styles, material aesthetics, and interactive mechanics through natural-language prompts before writing a single line of engine code.

The prototype is not the game: it is a way to find out whether the idea is worth building before the expensive work starts. Genie 4 is not a game engine, not a video generator, and not a validated robot simulator, and treating it as any of the three would be a mistake.

Genie 4 and Robots: The Apollo Checkers Demo

DeepMind's most concrete demonstration of Genie 4 outside entertainment is a robotics experiment, not a game. The team paired Gemini 4 Argon with Genie 4 to simulate Apptronik's Apollo humanoid robot playing a game of checkers, with Argon breaking the task into real-time steps that Genie then rendered as a simulated world. The steps were interleaved as text captions while the simulation rendered, and because Genie 4 can be steered in real time with text, the simulation could be adapted as it progressed toward completing the task.

Checkers is a deliberately small task, and that is the point. A robot that can be tested on a board game in a simulated world can, in principle, be tested on a warehouse aisle or a kitchen counter in the same way, without the cost or danger of the real thing.

Google's pitch is that world models give robots and agents a safe, infinitely variable training ground, and the Apollo demo is the first public evidence that the Genie and Gemini lines can work together toward it. Gemini 4 Argon itself became generally available on 13 October 2026, according to DataCamp's coverage, two weeks before the Genie 4 announcement.

Keep the scale in view, though. TechTarget's coverage of the announcement quotes analysts cautioning that sessions can drift and physics remain approximate, making Genie 4 unsuitable today for engineering environments that require precision.

A simulated checkers game proves the loop works; it does not prove the simulation is accurate enough to trust with a real robot. That gap between "the loop works" and "the physics are right" is the whole research problem DeepMind is working on.

How to Get Access to Genie 4: The Waitlist

There is no Genie 4 release date yet, and no general-availability date has been published at all. The only way in today is the waitlist, and the only people using the model are the trusted testers DeepMind started with.

The Genie 4 Project Genie rollout is the next step after the trusted-tester preview, and it is the route most readers will actually take. Google's rollout order is explicit: after the trusted testers come Google AI Ultra subscribers in Project Genie, and early-access partners of the Gemini API and Google AI Studio, both described as "coming soon" with no date attached.

1. Open DeepMind's Genie 4 announcement page and find the "Join the waitlist for Genie 4 in Project Genie" link.
2. Follow the link, which opens a Google-hosted application form for the waitlist.
3. Complete the form with the Google account you would use for Project Genie.
4. Wait for the rollout: Ultra subscribers in Project Genie and early-access API partners are next, with no date published.
5. While you wait, try Project Genie with Genie 3, which needs a Google AI Ultra subscription, a personal Google account, and an age of 18 or over, to learn the controls.

Buying Google AI Ultra today does not give you Genie 4 today. Project Genie still runs Genie 3, and Genie 4 for Ultra subscribers is "coming soon", so the subscription buys you the queue, not the model. That distinction is the single most useful fact in this whole guide if you are deciding whether to pay: the waitlist is free, the subscription is not, and neither one puts Genie 4 in your hands this week.

Geography is the one access question Google has not answered clearly. Project Genie launched in the United States on 29 January 2026 and was reported to expand to eligible AI Ultra subscribers worldwide in May 2026, with Google's plans page listing it in more than 140 countries as of late October 2026.

But some Google help pages still said US-only as of late October, so the safe statement is that Project Genie is reported as widely available while Google's own pages have not been fully synchronised. Whether the Genie 4 waitlist itself is US-only or global has not been stated anywhere, and this guide does not guess.

How Much Does Genie 4 Cost?

Genie 4 has no published price, because Google has published no per-token cost, no API model identifier, and no general-availability date for it. There is nothing to buy yet, and no price to compare. The only money in the picture is the subscription that gates the app it will eventually run in, Google AI Ultra, and that is a plan price, not a Genie 4 usage price.

The prices in the table come from Google's own plans pages as read by multiple outlets in late October and early November 2026: Ultra starts at $99.99 a month for five times Pro's usage, and the 20x tier is $199.99 a month. The multiples check out exactly: $99.99 divided by $19.99 is 5.0, $199.99 divided by $99.99 is 2.0, and $199.99 divided by $19.99 is 10.0, so the 20x tier costs ten times Pro, not twenty times the money. A year of the 5x tier is $1,199.88 and a year of the 20x tier is $2,399.88, both before tax.

One price tracker reports that Google's plans page has listed Project Genie as a 20x-tier feature only since 30 October 2026, which would mean the $99.99 tier no longer includes it, while other trackers still list it under both tiers. That conflict is unresolved in the sources this guide checked, so check the plans page yourself before you pay for access to Project Genie. Do not confuse the related Gemini 4 Argon pricing with Genie 4 either: Argon is reported at $2 per million input tokens and $10 per million output tokens through 30 November 2026, rising to $4 and $20 from 1 December, and that is a different model with a different price.

Project Genie vs Genie 4: The App and the Model

Project Genie is the app; Genie 4 is the model. That one sentence clears up most of the confusion around the announcement, because the two names sound interchangeable and they are not.

Genie 4 is Google DeepMind's fourth-generation world model, the engine that generates interactive worlds. Project Genie is the Google Labs experimental prototype, at labs.google/projectgenie, where Google AI Ultra subscribers create, explore, and remix AI-generated worlds in the browser. It is the consumer-facing product that runs a Genie model, and today it runs Genie 3, not Genie 4.

The timeline makes the split concrete. Genie 1 arrived in 2024, Genie 2 in December 2024, and Genie 3 was announced on 5 August 2025.

Project Genie launched on 29 January 2026 for AI Ultra subscribers in the United States, running Genie 3, and expanded to eligible subscribers worldwide in May 2026. Genie 4 was announced 271 days after Project Genie launched, on 27 October 2026, and it has not reached the app yet: Genie 4 is "coming soon" to Project Genie for AI Ultra subscribers.

Inside Project Genie, three Google models work together: Genie 3 runs the world, Nano Banana Pro creates the preview image you can edit before entering, and Gemini also powers the experience. The preview-image step is worth naming correctly, because a mistranscribed tutorial transcript has circulated a garbled name for it, and the model Google actually uses is Nano Banana Pro, its image-generation and editing model. Project Genie's Genie 3-era limits are a 60-second session cap, 720p at 24 frames per second, and a gallery for remixing other users' worlds, with Street View integration added in May 2026 for real United States locations.

What Genie 4 Cannot Do Yet

Genie 4 cannot hold a world together for long, and DeepMind says so in its own announcement. While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time, especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are all listed as active areas of ongoing research, not solved problems.

That list is the honest boundary of the model. A world whose details drift after a few prompt edits is fine for a filmmaker's draft or a game designer's prototype, where the point is to explore ideas quickly, and it is not fine for a robot-training run that needs the same doorway in the same place every time. The dev.to comparison of the two generations makes the same point from the other side: neither Genie 3 nor Genie 4 is currently a dependable substitute for a conventional simulation engine, and both remain research-stage systems with listed limitations across action space, multi-agent interactions, text rendering, spatial accuracy, and session duration.

The practical test is simple: if your use case breaks when the scene quietly changes behind your back, Genie 4 is not ready for it. If your use case survives a drifting detail, like a draft, a sketch, or a brainstorm, it is exactly what the trusted testers are using it for. Google's framing matches that test: the preview exists to learn alongside the community and to advance the frontier safely, not to ship a finished tool.

Is Genie 4 Safe? SynthID and Guardrails

Genie 4 includes strict safety guardrails consistent with Google's safety policies across other models, and DeepMind says it ran internal rigorous safety evaluations before launching the preview to prevent the generation of harmful or malicious content. The preview went to trusted tester partners first deliberately, to get early feedback before a broader release, rather than opening to everyone at once.

The transparency measure is SynthID, DeepMind's imperceptible digital watermark. All video and audio generated by Genie 4 will include SynthID, which helps identify AI-generated content across both visual and audio modalities.

That matters more for Genie 4 than for a still-image generator, because a world with native audio produces two kinds of output that could both be mistaken for real recordings, and the watermark covers both. The watermark is imperceptible, so it does not change what you see or hear; it changes what a detector can prove about where the clip came from.

If you want to try AI generation yourself today while Genie 4 is still behind its waitlist, our AI image generator is free to use in the browser, and our Ollama beginner guide shows how to run open models on your own machine. CodexFuse does not offer a Genie 4 tool, and no third-party site can offer you real Genie 4 access either: the only route in is Google's own waitlist, and any page claiming otherwise is not selling what it says it is.

"""Genie 4 numbers, recomputed 6 November 2026 (all dates from the sources below)."""
from datetime import date
genie3 = date(2025, 8, 5)
genie4 = date(2026, 10, 27)
project_genie = date(2026, 1, 29)
today = date(2026, 11, 6)
print((genie4 - genie3).days)          # 448 days between announcements
print((genie4 - project_genie).days)   # 271 days from Project Genie launch
print((today - genie4).days)           # 10 days since the announcement
print(120 / 60)                        # 2.0x: press-reported session vs app cap
print(round(99.99 / 19.99, 2))         # 5.0x: Ultra 5x vs Pro
print(round(199.99 / 99.99, 2))        # 2.0x: Ultra 20x vs Ultra 5x
print(round(199.99 / 19.99, 2))        # 10.0x: Ultra 20x vs Pro
print(round(99.99 * 12, 2))            # 1199.88: Ultra 5x annual cost
print(round(199.99 * 12, 2))           # 2399.88: Ultra 20x annual cost
print(1280 * 720)                      # 921600 pixels per 720p frame
print(1280 * 720 * 24)                 # 22118400 pixels per second at 24 fps
print(60 * 24, 120 * 24)               # 1440 frames in 60 s; 2880 in 120 s
# Expected output:
# 448
# 271
# 10
# 2.0
# 5.0
# 2.0
# 10.0
# 1199.88
# 2399.88
# 921600
# 22118400
# 1440 2880

Sources

Every fact in this article was checked against a source fetched on 6 November 2026. Primary sources are listed first.

• Google DeepMind, "Genie 4: A new generation of world simulation", 27 October 2026, deepmind.google — the announcement: four capabilities, native audio, Gemini 3.8 Flash, SynthID, limitations, waitlist.
• Google Blog, "Introducing Genie 4", 27 October 2026, blog.google — the Google blog version of the announcement.
• Google DeepMind, "Genie 3: A new frontier for world models", 5 August 2025, deepmind.google — Genie 3's announcement: 720p at 24 fps, a few minutes of consistency, promptable world events.
• TechCrunch, "Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering", 27 October 2026 — the two-minute session figure and the Ultra rollout in the coming weeks.
• TechTarget, "Google Genie 4 brings world models closer to robot and AI testing" — the Apollo checkers demo, analyst caution on drift and precision, no pricing or release dates.
• Indian Express, "What is Genie 4, Google's new world model that can simulate humans, physics, and sound?" — the announcement recap and use cases.
• i-scoop, "Genie 4 shows where world models are heading, and what European organisations can do now" — the research-preview status, no price or specs published, Project Genie not available in the EU.
• dev.to, "Genie 4 vs Genie 3: What Changed in Google's Interactive World Models?" — the independent comparison, including the limits both generations share.
• DataCamp, "Genie 4" — Gemini 4 Argon generally available 13 October 2026, and no Genie 4 specs published.
• Wikipedia, "Project Genie (website)" — Project Genie launched 29 January 2026, worldwide rollout May 2026, 60-second sessions.
• Apidog, "Project Genie free" — Project Genie requirements: Google AI Ultra, age 18 and over, personal Google account.
• AI Tool Analysis, "Google AI Ultra vs Pro 2026" — Ultra prices from Google's own plans pages, October 2026, and Project Genie listed as a 20x-tier feature since 30 October 2026.

Frequently Asked Questions

When will Genie 4 be available to the public?

There is no public date. Genie 4 was announced on 27 October 2026 as a research preview for trusted testers, and Google says it is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. The waitlist is open now through the link in DeepMind's announcement, and that is the only way to register interest.

Is Genie 4 free?

Genie 4 has no published price, because Google has not announced a price, an API model identifier, or a general-availability date. Joining the waitlist costs nothing. The app it will eventually run in, Project Genie, requires a Google AI Ultra subscription, which starts at $99.99 a month in the United States, but that is the price of the Ultra plan, not a price for Genie 4 itself.

What is the difference between Genie 4 and Project Genie?

Genie 4 is the model and Project Genie is the app. Genie 4 is Google DeepMind's fourth-generation world model, the engine that generates interactive worlds.

Project Genie is the Google Labs prototype where AI Ultra subscribers create and explore worlds in the browser, and it currently runs Genie 3. Genie 4 will reach Project Genie later, with no date given.

How long can a Genie 4 world run?

Press reports from the launch briefing say a Genie 4 scene can hold together for about two minutes, roughly double Genie 3's limit, but Google's own pages only say "a few minutes" and give no number. The 60-second cap people remember belonged to the Project Genie app, not to the Genie 3 model. Even at two minutes, details can drift, and DeepMind lists long-term consistency as an open research problem.

Can I use Genie 4 to make games?

You can use it to prototype a game idea in seconds, testing visual styles, material aesthetics, and interactive mechanics with text prompts before writing engine code. You cannot use it as a game engine, because it is a research preview that generates worlds on the fly rather than running rules a designer wrote, and its physics are approximate. For now it is a draft and ideation tool, not a production tool.