Google's Genie 4 Creates Worlds You Can Walk Through. Can You Actually Try It?

Imagine describing a landscape made of stained glass. Flowers rise from a luminous field, a river winds between mountains, and you move the camera through the scene as it comes into view. Then you ask for the weather to change while the world continues around you.

That is the sort of experience Google is exploring with Genie 4, an AI model designed to generate interactive environments in real time. The worlds can have sound, including ambient noise, music and dialogue.

For most people, Genie 4 is not available to try yet. Google has released it to trusted testers and plans to bring it to eligible Google AI Ultra subscribers through Project Genie. The existing Project Genie prototype is a separate experience that Google's help page still describes in terms of Genie 3.

Even so, Genie 4 gives us an intriguing look at what might come after the AI video clip: a scene you can move through and influence while it unfolds.

How do you actually explore it? For now, imagine navigating a world on your computer screen, much like moving through a video game. Google describes its demonstrations as using keyboard controls and text prompts to change the scene. A VR headset version could be an exciting future possibility, but Google has not announced one for Genie 4. Google DeepMind

A world that responds as you move

A world model is an AI system that generates a changing environment and predicts how that environment may respond to actions. Move forward, and it generates more of the scene. Ask for a change, and the system attempts to work that into what happens next.

Think of the difference between watching a filmed walk through a garden and choosing which path to take. The second experience needs a world that can keep responding as your choices change.

Google describes its demo recordings as showing users navigating Genie 4 scenes with keyboard controls and steering them with text prompts. Its examples include introducing a person or object, changing the weather and moving the camera. Google DeepMind

This is still a research simulation. A finished game usually has designed rules, objectives and environments that developers can test and refine. Genie 4 is exploring how much of that surrounding world an AI system can generate as the experience runs.

Sound and Gemini give the world more to work with

Genie 4 is Google's first real-time world model to generate audio natively, according to the company. It describes environmental sounds, music and dialogue that can respond to the scene. A generated beach could have waves as well as water; a generated character could have something to say.

Google has also brought Genie closer to models from its Gemini family. It says Gemini Omni helps anchor the appearance and behavior of a simulation. In the system Google describes, Gemini 3.8 Flash observes the scene, breaks goals into steps and checks progress as the world develops. Google DeepMind

One way to picture the relationship is that Genie produces the immediate sights and sounds while Gemini helps guide the unfolding sequence. Google says Gemini also adapts the user's prompts to fit the surrounding scene.

Genie 3 already opened the door

Genie 4 follows a model that could already generate navigable worlds. Google's 2025 Genie 3 announcement also showed prompted changes to an environment, so the idea of interacting with a generated scene did not begin this week.

The new version adds native audio and places more emphasis on Gemini-guided events. Google also claims improvements in visual realism and human movement. Together, those changes point toward scenes that feel more like places with events happening inside them.

The product distinction matters here. Genie 3 is a model; Project Genie is the experimental interface Google built for people to create and explore worlds. At launch, Google said Project Genie combined Genie 3 with Nano Banana Pro and Gemini. Google's Project Genie announcement

Today’s Project Genie runs Genie 3

Google's Project Genie help page describes the existing prototype as available to eligible Google AI Ultra subscribers where the plan is offered. Users must be at least 18 and sign in with a personal Google account.

The documented experience lets people create a world from text or images and explore it for 60 seconds. Google says generation currently uses no AI credits during this research prototype. Its launch post says promptable events that change the world during exploration are not included in the prototype; the current Help page documents navigation controls. Street View location grounding is limited to US locations. Account access and availability can vary. Google’s launch post

Genie 4 is the next step. Google says it is coming to Ultra subscribers through Project Genie and to early-access partners of the Gemini API and Google AI Studio. TechCrunch reports that the rollout is expected in the coming weeks. Google has not provided a firm date in its announcement.

Google’s announcement links a Genie 4 waitlist. The future experience's precise subscription conditions, session limits and regional rollout should be checked when it arrives.

The demonstrations show promise and leave questions

Google's examples are selected demonstrations of the system it is researching. Choosely has not used Genie 4, so we cannot say how consistently ordinary users will reproduce those results.

Google acknowledges that objects and details can drift as sessions go on, especially after complex edits or movement through a scene. It lists complex physical interactions, audio accuracy, legible text, spatial accuracy and long-term consistency as areas that still need work. Google DeepMind

Google describes coherence over short sessions in its announcement. TechCrunch reports that a scene can remain consistent for up to two minutes; that figure should not be treated as a confirmed limit for the eventual consumer product.

This matters for the applications being explored. A beautiful image may be useful for inspiration even when an object shifts unexpectedly. A simulation used to evaluate a robot needs a much stronger account of how faithfully actions and physical consequences are represented.

Creators can explore ideas; researchers can explore scenarios

For a filmmaker, a navigable scene could offer a fresh way to explore a setting, mood or camera direction during early ideation. TechCrunch's reporting describes trusted testers exploring film prototyping and quotes Google on storyboarding and rapid creative exploration. How well that will fit an ordinary production workflow remains to be seen.

Researchers have another reason to pay attention. Google describes an early experiment in which Gemini 4 Argon creates a multi-step plan that Genie 4 simulates for robotics research. A wider range of generated situations could give researchers more scenarios to examine before taking systems into the physical world. Google DeepMind

And there is a straightforward appeal for everyone else: being able to describe a place, enter it and see where a little curiosity takes you.

Choosely's take

Genie 4 makes the next direction of generative AI easier to picture. Sound, scene steering and Gemini's guiding role give Google's interactive worlds more room to develop into a creative medium.

Don't upgrade to Ultra solely for Genie 4 yet. It remains a restricted preview, and the consumer rollout and its conditions need confirmation. If you already have Ultra and the idea appeals, join Google's waitlist and explore the current Project Genie prototype if your account is eligible.

The next useful milestone will be independent access. Once people can test their own prompts and explore beyond Google's chosen examples, we'll learn much more about how these worlds behave.