Genie 4 transforms AI worlds into interactive spaces
Google DeepMind introduces Genie 4, a model capable of generating 3D worlds, sound, and real-time dialogue. Access begins in research preview, before gradual opening to Gemini API, Google AI Studio, and AI Ultra.
Google DeepMind has unveiled Genie 4, a world model capable of generating interactive three-dimensional environments, with realistic visuals, native sound, and real-time dialogue. Presented as a research preview, it is reserved for trusted testers in the creative, journalistic, educational, video game, and industrial sectors. The challenge goes beyond a technical demonstration: Google aims to make these simulated worlds a testing ground for robotics, AI agent training, and interactive content production.
Worlds that react, not just to be viewed
Genie 4 is distinguished first by its ability to maintain a coherent scene while the user interacts with it. Testers can modify the setting by text, change the weather, move the camera, or add objects and characters without interrupting the simulation. The model also generates soundscapes and interactive dialogues; according to Google DeepMind, this is the first time Genie produces audio natively in real time.
This combination places the system halfway between a game engine, a prototyping tool, and a sandbox for agents. Google states that Genie 4 can be used to train AI systems to navigate complex environments before their deployment in the physical world. The word “simulation” takes on a very concrete meaning here, and is a little less dusty than usual.
Google also pushes usage towards creation and robotics
The company cites several targeted uses: game development, film staging, exploration of historical scenes from archival photographs, or visualization of events such as sudden floods and forest fires. In a demonstration, an Apptronik humanoid robot Apollo plays chess in a generated environment, while Gemini 4 Argon breaks down the task into steps and monitors its execution. This duo shows Google's ambition: to link a simulated world to a decision chain for robotic systems.
Genie 4 relies on several components from the Gemini family. Gemini Omni provides a base for the appearance and behavior of the world, while other models handle reasoning to maintain simulation coherence. According to the announcement, Gemini 3.8 Flash acts as a guiding agent in complex scenarios: it observes the environment, breaks down tasks, tracks progress, and readjusts the plan if the user intervenes or if the setting changes.
Limited access, then gradual opening
At this stage, Genie 4 remains a research product. Access is limited to trusted testers, but Google plans to expand availability to Google AI Ultra subscribers via Project Genie, as well as to certain partners with early access via the Gemini API and Google AI Studio. This scaling up will be decisive: it will determine whether Genie remains a spectacular lab or establishes itself as a tool usable in real workflows.
Google DeepMind also states that SynthID is applied to all video and audio generated by Genie 4. This digital watermark allows identification of AI-produced content, in both visual and audio formats. At a time when fake content spreads quickly, this safeguard does not solve everything, but at least prevents generated worlds from pretending to have been born from a shoot in the rain.
Key points
- Genie 4 generates interactive 3D worlds in real time.
- The model produces native audio and interactive dialogues.
- Initial access reserved for trusted testers.
- Google AI Ultra and Gemini API will follow via Project Genie.
- SynthID will be applied to audio and video content.
- Apptronik and Gemini 4 Argon appear in a robotic demo.
By the numbers
- 4 — Genie 4, new generation announced by Google DeepMind.
- 3D — three-dimensional environments generated in real time.
- 2026 — publication date of the secondary source citing the launch.
- 2 — main access families planned: AI Ultra and API partners.