Technology•

Google Opens Genie 4 to AI Ultra Subscribers, With Worlds That Can Be Rewritten Mid-Exploration and Characters That Talk Back

Google began rolling out Genie 4 to eligible AI Ultra subscribers on November 18 through Project Genie, adding mid-session text edits, generated audio and character dialogue, plus longer Genie 3 sessions.

Written By :MINRK
Google Opens Genie 4 to AI Ultra Subscribers, With Worlds That Can Be Rewritten Mid-Exploration and Characters That Talk Back

Photo by Frantzou Fleurine on Unsplash

1. Genie 4 reaches paying subscribers

Google began rolling out Genie 4, its newest world-generating artificial intelligence model, to eligible Google AI Ultra subscribers on Wednesday, November 18, 2026. Delivered through the browser-based Project Genie experiment, the model lets users build explorable environments, alter them with typed prompts while moving through them, and fill them with characters that speak. The Ultra plan costs $199.99 per month in the United States.

The release ends a period in which the model was visible but largely out of reach. Google first showed Genie 4 in October, yet hands-on access at that stage was restricted to a group of trusted testers. Wednesday's announcement moves it into the hands of ordinary paying customers for the first time, although only those on the company's most expensive consumer AI tier.

Subscribers one level down are being asked to wait. Members of the AI Pro plan can add their names to a waitlist, and Google has not given a date for when that group will be admitted. For now, the practical effect is that Genie 4 functions as a premium feature that helps distinguish the Ultra subscription from cheaper plans.

The model did not arrive alone. Alongside it, Project Genie gained an AI creation assistant, a community gallery for making and sharing worlds, and longer sessions for the older Genie 3 model, which remains available next to its successor.

2. What is Genie 4?

Genie 4 is the latest model in Google's Genie line, a family of AI systems that produce interactive environments a person can move through, generating each scene as the exploration happens rather than rendering a finished video from a prompt. That distinction is the core of the product, and it is what separates Genie from the text-to-video generators that have become familiar over the past few years.

A video generator takes a description and returns a clip. The clip may look convincing, but it is fixed: the viewer watches it and has no ability to turn left, open a door or walk toward something in the distance. Genie works differently. The model responds to what the user does, producing the next moments of the scene on the fly, so the output is closer to a place than to a recording.

What Genie 4 adds is the ability to redirect that place while it is running. Partway through a session, a user can type an instruction to change the weather, introduce a new character or steer what happens next, and the model folds the request into the world without requiring a restart. In earlier versions of this kind of tool, the description at the start largely determined the experience. Here the prompt becomes an ongoing control, something closer to directing a scene than to commissioning one.

Systems of this type are often described as world models, because their task is to maintain a plausible, continuous environment and predict how it should evolve in response to actions. The appeal is broad. A model that can simulate a space on demand could serve entertainment, creative prototyping and education, and the same underlying capability is of interest to researchers who want simulated settings in which to train or test AI agents.

3. Sound, speech and physical behavior

Audio is the other headline change. Genie 4 produces sound together with its visuals, and Google's description of the feature covers ambient effects such as wind and footsteps as well as dialogue. The speech is not a fixed soundtrack laid over the pictures. According to the company, characters respond to what is happening in the scene, which is what allows a user to populate a world with figures that answer back.

Google also says the new model handles motion more convincingly than its predecessors, pointing to improvements in the way people move and in the way objects react to physical forces. These are precisely the areas where generated environments tend to break the illusion, since a figure that walks unnaturally or an object that ignores gravity is immediately noticeable to a human observer.

Those claims come with an important qualification. They rest on demonstrations that Google itself produced, and they have not been confirmed through independent testing. Until outside users and researchers have spent time with the system, the degree of improvement in realism remains the company's own assessment rather than an established result.

4. From side-scrolling experiments to explorable 3D spaces

The distance the project has covered becomes clear when the current release is set against its origins. In 2024, Google presented the first Genie as a research effort centered on side-scrolling experiments. That initial version learned from more than 200,000 hours of gameplay footage and could take a single image and convert it into a simple two-dimensional environment that a person could play.

Later that same year, Genie 2 carried the research into three dimensions. Its generated environments could be navigated with a keyboard and mouse, either by people or by AI agents, which shifted the work from flat, platform-style scenes toward spaces with depth.

Genie 3 pushed in the direction of continuous exploration in real time, and it became the foundation of the first public product. Project Genie opened in January to adult AI Ultra subscribers in the United States, giving people outside Google their first sustained opportunity to try the technology in a browser.

That early experience was limited. Sessions ran for only 60 seconds, and both character controls and the physical behavior of the generated worlds could be unreliable. The updates announced this week are in large part a response to those constraints, addressing the length of sessions on one model and the quality of visuals and sound on the other.

5. Genie 3 stays available, with longer sessions

Rather than retiring the previous model, Google is keeping Genie 3 in service and extending what it can do. Sessions on Genie 3 now last three minutes, a threefold increase over the 60-second limit that applied when Project Genie launched. The older model also gains support for 360-degree panorama images and the ability to link several generated worlds together into longer journeys.

The three-minute figure applies specifically to the updated Genie 3 experience. Google did not state a comparable session length for Genie 4, leaving open how long a user can remain inside one of the newer model's environments.

The result is a choice between two sets of strengths. Users who want the stronger visuals and the generated audio will select Genie 4, while those who value more time inside a world, or who want to chain environments into an extended sequence, have reason to stay with Genie 3. Offering both side by side suggests that the capabilities have not yet been combined in a single model.

6. How is Genie different from a game engine?

Genie does not build a stable level governed by fully defined rules, which is what a traditional game engine does. Instead, the model predicts what a world should look and sound like as it changes, generating each new moment from its understanding of what came before and what the user has just done.

The difference has practical consequences. In a conventional game, the geometry of a room, the position of an object and the laws of motion are stored explicitly, so a chair that sits in one corner will still be there when the player turns around. A predictive model holds no such fixed record. It produces a plausible continuation, and plausibility is not the same as consistency.

Google is open about the limitations that follow. The company acknowledges that objects can drift or change between views, that complex physics remain difficult for the system, and that generated speech or on-screen text may be inaccurate. Those caveats place Genie 4 firmly in the category of an experiment, as the Project Genie label indicates, rather than a replacement for the tools that game studios use to build finished products.

The trade-off runs in both directions. A game engine demands that every asset and rule be created in advance, typically by teams of specialists, whereas a generative world can be summoned from a sentence and changed with another. What Genie gives up in reliability, it gains in speed and flexibility of creation.

7. Watermarks, safety screening and community tools

Because the model produces realistic imagery and synthetic voices, Google has attached safeguards to its output. The company says the audio and video generated in Project Genie carry SynthID watermarks, its technology for marking AI-made content so that it can later be identified as machine-generated. Google also says that what users put into the system and what the system returns both pass through safety screening.

The remaining additions concern how people create and share. An AI creation assistant has been added to help users build their worlds, and a community gallery gives them a place to create and share what they have made. Together with longer Genie 3 sessions and linked worlds, these features move Project Genie beyond a solitary demonstration and toward something users can build on and return to.

8. What comes next for Genie 4

Several questions remain open after the rollout. Google has not said when Genie 4 will reach AI Pro subscribers, who for now have only the waitlist. It has not specified how long a Genie 4 session lasts. And its claims about more natural human movement and more believable physical interactions still await examination by people outside the company.

For the moment, access is defined by price and eligibility: Genie 4 is available to eligible AI Ultra subscribers paying $199.99 per month in the United States, with Genie 3 running alongside it for those who prefer longer explorations. How quickly the model moves down to less expensive plans, and how well it performs once a wider group of users begins testing its limits, will determine whether real-time world generation becomes a mainstream feature or stays a showcase for Google's top subscription tier.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY