Genie 4 waitlist: how to try Google’s AI worlds, and the catch

Google opened the waitlist alongside the announcement, and it’s handled through Project Genie, the same place people have used the older Genie 3 model.

1. Go to the Project Genie site and sign in with your Google account.
2. Apply for Genie 4 access through the waitlist form.
3. Watch for an invitation. Google hasn’t given a firm timeline for wider access.

Joining costs nothing. Being a paying Ultra subscriber is what’s expected to move you up, since that plan is first in line once testing widens. AI Pro and AI Plus subscribers may follow later, but Google hasn’t committed to it.

Access comes in stages:

• Now: trusted testers only. DeepMind says these groups work in creative industries, journalism, education, gaming and engineering.
• Soon: Google AI Ultra subscribers, inside Project Genie.
• Later: early access partners using the Gemini API and Google AI Studio, Google’s tools for developers. Developer pricing hasn’t been announced.

For now, it runs only on Google’s servers. There’s nothing to download, and you won’t be running it on your own gaming PC.

Genie 4 is a world model. Instead of making a fixed video clip, it builds a scene frame by frame as you move through it, reacting to what you do. You describe a place in text, or give it an image, and then walk around with keyboard controls.

Four upgrades stand out over Genie 3, which Google showed in August 2025:

• Sound. It’s DeepMind’s first real-time world model that makes audio itself. Each world has its own ambient sound, from wind in trees to waves on a beach, and characters can hold conversations that react to you.
• Longer worlds. Scenes stay consistent for about two minutes, double the one minute Genie 3 managed.
• Live changes. You can type a request mid-session, like new weather, a different camera angle or an extra person, and the world adapts without starting over.
• Better people and physics. DeepMind calls human movement its biggest leap, and says it handles collisions and aerodynamics more convincingly.

Behind the scenes, Gemini 3.8 Flash acts as a guide. It breaks your goal into steps, checks each new frame to confirm a step worked, and changes its plan if the world drifts or you interrupt. Another Gemini model keeps the look and behavior of the world grounded in reality.

DeepMind sees this as an early step toward what it calls hybrid world models. In that setup, a language model like Gemini plans the structure and story of a scene in the background, while Genie handles what you actually see and hear. Google also links the work to its long-term goal of artificial general intelligence, AI that can handle most tasks a person can, because understanding how the physical world behaves is a big piece of that.

The limits matter if you’re thinking of paying $100 a month for this. Google admits that details and object permanence, meaning things staying where you left them, start to slip in longer sessions, after complicated edits or while moving around. Interactions last a few minutes, not hours.

DeepMind’s own model page lists more open problems:

• Real places aren’t simulated with perfect accuracy.
• Readable text usually only appears when you ask for it in your description.
• Real-time audio accuracy still needs work.
• Worlds with several characters interacting at once are unsolved.

Google also hasn’t shared the usual specs, such as resolution, frame rate or test scores. Genie 3 ran at 720p and 24 frames per second, but there’s no equivalent figure for the new model yet.

When Genie 3 appeared, gaming stocks fell, because the output looked a lot like a video game made without a game engine. Genie 4 will revive that worry. Google pitches it for prototyping game ideas, though, not replacing finished games, and two-minute worlds aren’t a threat to a big adventure game you play for weeks.

The early uses are telling. Filmmakers are running several live takes of a scene to plan shots. Photojournalists are testing how one archival photo can become a walkable simulation. Game developers are trying out visual styles and mechanics before writing engine code. Robotics teams want endless practice worlds for robots, and in one demo Genie 4 simulated a humanoid robot playing checkers.

For everyday viewers, the useful detail is labeling. Every video and audio clip Genie 4 makes will carry SynthID, Google’s invisible watermark for spotting AI-made content. That matters as clips from tools like this start turning up in social feeds.