On October 27, 2026, Google DeepMind announced Genie 4, its new "world model." A world model is an AI that uses text or images to create, in real time, a virtual world you can walk around in and interact with. Genie 4 makes its visuals far more realistic and can now also generate ambient sound, music and character dialogue at the same time. It is currently offered to a limited group of testers as a research preview.
What is Genie 4?
In Genie 4, when you make something happen, the AI goes on to generate how the world responds, what it sounds like, and how the people in it move and speak. Users control movement and the camera direction with the keyboard, and change the scene with text.
- Changing the world with words: users can give instructions on the spot, such as changing the weather, moving the camera or bringing in new characters and objects
- Reproducing physics: it creates motion close to real-world physics, such as airflow and objects colliding
- Character movement: people move naturally and react to changes around them. Google DeepMind calls this its biggest advance
- Sound generation: it simultaneously creates sounds specific to each place, such as wind rustling the trees or waves, as well as music. Genie 4 is the company's first real-time world model that generates sound along with the world
- Dialogue: characters speak on the spot in response to the user's actions and events in the world
Google DeepMind sees simulating the world as an important step toward AI that can handle as broad a range of tasks as humans (AGI). It says the model can help build environments for safely training robots and AI agents, and support production work in fields such as film, journalism, games and manufacturing.
Working with Gemini
Genie 4 draws on the capabilities of Gemini models. Gemini Omni makes the world's look and motion closer to reality, while other Gemini models handle the reasoning that keeps the world consistent.
For example, Gemini 3.8 Flash acts as a guide that keeps watching the world as it is generated. It breaks a goal into several steps and checks the video to confirm whether each step succeeded. If the user changes the instructions midway, it revises the plan. Google DeepMind also showed an example in which Gemini 4 Argon and Genie 4 are combined to simulate, step by step, Apptronik's humanoid robot Apollo playing checkers.
How testers are using it
Google DeepMind has been testing the model with testers in film and video production, journalism, education, games and engineering. The uses it highlighted are as follows.
- Film and video production: stepping into a scene before shooting and trying take after take while changing camera angles, lighting and lines
- Journalism and history: recreating a place as a living world in seconds from a single old photograph. Testers are also trying it to show how floods and wildfires spread
- Games: turning an idea into an explorable prototype in seconds, and testing its look and gameplay with words
- Simulation: companies are evaluating it as a tool for quickly creating a wide range of scenarios
Availability plans
Genie 4 is first being offered to a limited group of testers. Google DeepMind plans to make it available through Project Genie for Google AI Ultra subscribers, and to participants in early access for the Gemini API and Google AI Studio. The timing has only been described as "soon." Project Genie is accepting waitlist sign-ups. The announcement does not give pricing or the regions where it will be offered, including Japan.
Safety measures and remaining challenges
Google DeepMind says it has put in place restrictions based on the same safety policies as its other models, and began the preview after internal safety evaluations. The video and audio Genie 4 creates carry SynthID, an invisible watermark that identifies them as AI-generated.
At the same time, the company describes Genie 4 as still an early, experimental model. It can keep the world consistent over short periods, but when it runs for a long time, receives many layered complex instructions, or the user moves around within the scene, details and the positions of objects can drift. Complex physical interactions, audio accuracy, rendering text, spatial accuracy and long-term consistency are all still being researched.
CoAI's take
World models are a field where companies are racing to develop the technology. On October 24, Tencent also released Hy-Nomas, an open world model. What sets Genie 4 apart is that it generates not just video but also sound and dialogue on the fly, and works with Gemini to keep events unfolding coherently.
However, it will be a while before general users can try it. The first to get access will be subscribers to Google AI Ultra, the top-tier plan, and some developers, and it is not known whether it will be available in Japan. Those interested in video production or game prototyping should gauge its capabilities from the demo videos while waiting for announcements on regions and pricing.
