Genie 4: what it is and how to get access
2026-10-27 · 4 min read
Genie 4 is Google DeepMind's new world model: you describe a scene, it generates a world you can walk through in real time, with sound, and you can change that world mid-session by typing. Right now you get access only as a trusted tester, with a rollout to Google AI Ultra subscribers in Project Genie and to early access partners on the Gemini API and AI Studio promised "in the coming weeks."
@GoogleDeepMind announced it on October 27 with a two-and-a-half-minute demo reel, calling it "our most realistic and interactive world model yet." The post says you can see how a scene "responds, sounds," and how the people in it move and speak. The audio part is new for Genie.
What changed from Genie 3
Genie 3 came out as a research preview a little over a year ago. The biggest addition is sound: music, background noise and character dialogue are generated along with the world, and DeepMind says the characters talk back in real time based on what you do.
You can also change a scene while it runs. Type a line to make it rain, move the camera, or add a person or an object, and walk around with normal keyboard controls. TechCrunch reports that a scene now stays consistent for up to two minutes, twice Genie 3's limit.
Under the hood, Gemini now sits in the loop. DeepMind says Gemini 3.8 Flash acts as a guiding agent during generation. It watches the scene, splits your goal into steps, checks each step, and changes its plan when you step in. Gemini Omni handles making things look and behave like the real world. DeepMind's name for this is "hybrid world models": Gemini plans the story and Genie draws it.
What Genie 4 can do
| Area | What DeepMind shows or reports |
|---|---|
| Scenes from text | A saxophonist on a floating platform in a misty lake, a 1980s rural Japanese train station, a stained-glass landscape you can walk through |
| Physics | Aerodynamics, collisions, a maple seed falling in macro view |
| People | Characters who move, react and talk back in real time |
| Robotics | An experiment where Gemini 4 Argon split a task into steps for Genie 4 to simulate, with an Apptronik Apollo humanoid playing checkers |
| Tester uses | Film previsualization, turning archival photos into scenes, prototyping game styles before writing engine code, scenario generation for enterprises |
| Safeguards | SynthID watermark on all generated video and audio |
DeepMind is open about the weak spots. Details and objects can drift on longer runs or after many edits, and in-world text, audio accuracy and spatial accuracy are still being worked on. Google engineer Diego Rivas told TechCrunch it is "still early days for Genie 4."
How to get access
There are three routes in right now:
- Join the waitlist. DeepMind points interested users to one through Project Genie.
- If you already pay for AI Ultra, watch Project Genie over the next few weeks. Google's subscriptions page lists Project Genie on the $200 a month Ultra tier, for subscribers 18 and older. The cheaper $100 Ultra tier doesn't include it.
- If you're a developer, API access goes first to early access partners in the Gemini API and AI Studio. Google hasn't published API docs or per-use pricing.
When you're in, start with the kind of prompts DeepMind demoed: a specific place, a camera position and a mood. "First-person view from a hot air balloon over Antarctica" is one of theirs. Then test the steering, since that's what's new. Make it snow, add a character and talk to them, swing the camera around, and see how long the scene holds together before things start to drift.
The competition
Google isn't the only lab doing this. On September 3, @runwayml posted its GWM Worlds 2, "our latest General World Model that generates interactive real-time simulations," which also takes text actions and has no preset session length. Runway's pitch is open-ended sessions. Genie 4's is DeepMind's physics, the Gemini planner, and characters you can talk to.
My read
I wouldn't plan any business work around Genie 4 this quarter. With a two-minute ceiling and drift, it can't do anything that needs the same reliable output twice. The tester uses are still worth reading, because a few of them fit a small shop once access opens: storyboarding a video shoot, mocking up a game idea before you pay a developer, maybe walking a client through a space that hasn't been built. For browser games specifically, Google Labs Playground already lets you try a version of that.
I think the robotics experiment will matter more in the long run. Google already pitches its models for physical work, as with Gemini Robotics 2, and a model that can rehearse a task before a robot tries it would make that work safer. For now, Genie 4 sits behind Google's most expensive plan. If you're comparing plans, our breakdown of who gets Gemini 4 Argon explains how Google splits its tiers.
If you'd like to know which parts of your own week AI can handle today, New Face Design's free process audit is a good place to start. We'll go through your workflows with you and point out the few worth automating first.