Genie 4 is Google DeepMind’s new world model, and the big change is sound. On 27 October 2026 DeepMind showed a system that builds a playable world from a prompt and, for the first time, generates its music, ambient noise and character dialogue live as you move through it.
It is a research preview, not a product you can download. Still, this is the model family that knocked billions off video game stocks earlier in 2026, so it deserves a careful look.
Key takeaways
- Genie 4 was announced on 27 October 2026 as a research preview for trusted testers.
- It is DeepMind’s first real-time world model to generate audio natively, including speech.
- You can edit a running world with plain text: change the weather, move the camera, add people.
- According to TechCrunch, scenes stay consistent for up to two minutes, double Genie 3’s limit.
- Google AI Ultra subscribers in Project Genie and some Gemini API partners are next in line. No public date or price yet.
What is Genie 4, in plain terms?
A world model is an AI that predicts what the next moment of a scene should look like, based on what you just did. Press forward, and it draws the road ahead. Turn left, and it invents what is around the corner.
Think of it as a game engine with no code inside. A normal game stores a level as files. Genie dreams the level frame by frame, the way you might picture a room you have never visited.
In its official announcement, DeepMind says Genie 4 “delivers enhanced photorealism, fluid movement, and native audio.” It also claims better physics, such as aerodynamics and collisions, and characters that react to their surroundings.
What changed from Genie 3 to Genie 4?
Genie has moved fast. The first version in March 2024 made 2D platformer scenes. Here is how the line has grown, based on DeepMind’s posts and reporting from TechCrunch and Wikipedia’s Genie history.
| Version | Announced | What it could do | Scene memory |
|---|---|---|---|
| Genie 1 | March 2024 | 2D worlds, about 1 frame per second | Not stated |
| Genie 2 | December 2024 | 3D worlds at 360p | 10 to 20 seconds |
| Genie 3 | August 2025 | Real-time 3D at 720p, 24 fps, no sound | About 1 minute |
| Genie 4 | 27 October 2026 | Photoreal worlds, native audio, live text edits, dialogue | Up to 2 minutes (TechCrunch) |
DeepMind has not published Genie 4’s resolution or frame rate, so treat any figures you see elsewhere with care.
How does Gemini fit into Genie 4?
Genie draws the world. Gemini does the thinking. DeepMind describes Gemini 3.8 Flash as a guiding agent that plans steps, checks progress and adapts when the world changes or you step in.
Google research director Tejas Kulkarni told TechCrunch: “What we’ll increasingly see in the future are these hybrid world models.” In a robotics experiment, a model called Gemini 4 Argon broke tasks into steps that Genie 4 then simulated in real time.
That robotics angle matters more than it sounds. A robot that can rehearse a task in a simulated kitchen, thousands of times, is cheaper and safer than one learning by breaking real plates.
Why did Genie rattle video game stocks?
When Project Genie, the consumer site running Genie 3, rolled out more widely in late January 2026, investors panicked. As Startup Fortune recounts, Unity Software fell more than 20% in one session, Roblox dropped about 13% and Take-Two lost roughly 8% to 10%. Bloomberg’s headline on the Genie 4 launch leans on that same history.
The fear was simple: if anyone can type a game into existence, who needs engines and studios? Unity CEO Matthew Bromberg pushed back, arguing that the output is not suited on its own to games that need consistent, repeatable play.
My view, clearly as opinion: he has a point. A two-minute world that can drift is a wonderful sketchpad, not a shipping game. Games need saves, rules and levels that look the same tomorrow. Genie 4 narrows the gap; it does not close it.
What Genie 4 still gets wrong
DeepMind is unusually frank about the limits. It lists these as open problems:
- Drift: visual details and object permanence (whether things stay where you left them) can shift over longer runs or after several edits.
- Audio accuracy: sound does not always match the scene.
- Text: signs and writing inside the world are often unreadable.
- Spatial accuracy and long-term consistency: still active research.
Every generated frame and sound carries SynthID, Google’s invisible watermark, so clips can be traced back to the model later.
What happens next
Trusted testers have access first. TechCrunch reports that over the following weeks, Genie 4 should reach Google AI Ultra subscribers inside Project Genie and early access partners of the Gemini API and Google AI Studio. A waitlist is open. Google has not named a public release date, countries or developer pricing.
The competition is heating up too. Runway shipped GWM Worlds 2 a couple of months earlier, and TechCrunch notes AMD is acquiring World Labs. If you follow AI video, our look at xAI’s Grok Odyssey film and the EVEngine AI platformer demo show other routes to the same goal.
FAQ
Can I use Genie 4 now?
Not unless you are a trusted tester. Google AI Ultra subscribers in Project Genie and selected Gemini API and AI Studio partners are next. You can join the waitlist.
Does Genie 4 make real video games?
No. It generates short, playable worlds that can drift over time. It is useful for prototyping, learning and robot training, but it does not produce a finished game with fixed rules and assets.
What is new in Genie 4 compared with Genie 3?
Native audio and dialogue, live text edits mid-session, better physics, more lifelike characters and, per TechCrunch, scenes that hold together for up to two minutes instead of about one.
Want a turn when the doors open? Join the Project Genie waitlist, and keep your first prompt to one simple, vivid scene.
Leave a comment