Genie 4: Google Constructs Worlds You Can Walk Into. And They Talk Back.
You type a sentence. You don't get a video, but a place you can walk through yourself.
Figure 01.0 — Genie 4
First things first: what is a world model?
Most AI video tools create a clip. You watch it and that's it.
A world model works differently. It creates an environment that responds to your actions. You navigate with your keyboard, just like in a game. Only, there is no game. The AI invents the world as you walk.
What is it?
Genie 4 is the latest world model from Google DeepMind. Announced on 27 October 2026, it follows Genie 3. blog
You can initiate an event and direct the scene using ordinary language. Move the camera, change the weather, add a character. blog
[IMAGE: demo from the Google DeepMind blog, such as the train station]
What's new?
Sound. According to DeepMind, this is their first real-time world model that generates audio itself. It includes ambient sounds, music, and dialogue, all generated on the spot. ppcblog
Characters can also engage in conversations that respond to your actions. DeepMind calls this the real boundary of this release. martincid
The demos showcase its breadth. A maple seed drifting to the grass. A bustling rural station in Japan, 1980s. A dragon race from the rider's perspective. All sound is generated by Genie 4 itself. martincid
Under the hood, Genie 4 collaborates with Gemini. Gemini Omni ensures it looks real, while other Gemini models keep the scene logical and coherent. blog
Where it still falters
Google is candid about it. Objects that persist when you look away, and readable text, remain weak points. DeepMind describes it as an early, experimental model. datacampppc
There are also no hard numbers yet. Google has not published benchmarks or comparisons, and independent tests have not been conducted. datacamp
Everything created by Genie 4 is marked with SynthID, Google's watermark, so it can later be identified as AI. cellcog
A quick look back
It progressed rapidly. In early 2024, Genie was still creating simple 2D platform games from an image. By December 2024, 3D arrived with Genie 2. In August 2025, Genie 3 allowed real-time navigation. Now, sound and voices have been added. martincid
Investors are also watching. Bloomberg reports on a world model that unsettled shares of gaming companies. Bloomberg
Can I try it?
Not yet. A small group of testers currently has access. Following this, Google AI Ultra subscribers via Project Genie, and early partners via the Gemini API and AI Studio will gain access. There is a waiting list, but no set date. blogthenextweb
With the previous version, it took a long time. On X, the most common complaint was that Genie 3 was still not available to all Pro subscribers a year later. datacamp
To follow
When the waiting list opens. And what creators will do with it. Photojournalists have already attempted to convert archival photos into simulations. This might be the most intriguing aspect of this entire story. thenextweb
