Google DeepMind opens Genie 4 world model to trusted testers

Google DeepMind released Genie 4, its newest world model, as a research preview for trusted testers on Tuesday, October 27. A world model generates an interactive environment frame by frame as you move through it, with no pre-built levels or hand-programmed objects. ¹

The headline additions are native audio, free-form steering and better physics. Genie 4 generates music, ambient noise and interactive dialogue alongside the picture. ¹² Users can type prompts mid-simulation to change the weather or camera angle, or to add people and objects. ¹ DeepMind says aerodynamics and collisions look more realistic, and that people in the worlds move and react more naturally. ³

Gemini is more tightly woven in. Gemini Omni anchors the simulation in real-world appearance and behavior, while other Gemini models handle the reasoning that keeps a scene coherent. ² Ars Technica and The Verge both name Gemini 3.8 Flash as the model working under the hood. ³⁴

The limits are stated plainly. A scene holds together for up to two minutes, double Genie 3’s limit. ¹⁴ Details and objects can drift the longer a simulation runs, and DeepMind has not yet worked out how to guarantee accurate audio or legible in-world text. ¹ A DeepMind engineer stressed it is still early and that the end-to-end product experience is unclear. ¹

Access will widen soon. Google AI Ultra subscribers get Genie 4 through Project Genie, along with early access partners of the Gemini API and Google AI Studio, over the coming weeks. ¹ There is a waitlist for Project Genie. ² Ars Technica reports Ultra plans start at $100 per month and that no developer pricing or firm timeline has been announced. ³

Where do games fit? Google’s pitched uses are gaming, interactive learning and robotics. ¹ The Verge says Google lists prototyping game ideas alongside film pre-visualization and robot training scenarios. ⁴ Testers are pointing to film prototyping, with a DeepMind engineer saying real-time generation matters most in ideation and storyboarding. ¹ The robotics case is the most concrete demo: in one experiment, a Gemini model broke tasks into steps that Genie 4 simulated in real time, a possible way to test robot scenarios before they happen physically. ¹

Genie 3, announced in August 2025, generated worlds at 24 frames per second at 720p and stayed consistent for a few minutes. ⁵ Genie 4 does not appear to change what a world model is, but it fills in the pieces that make a generated scene feel less like a silent clip.

Our read: Two minutes of stable, steerable world with sound is a prototyping and research tool, not a game engine. Nothing in the preview suggests persistent levels, save states or the content pipeline a shipped game needs. The near-term fit looks like idea sketching and agent training. Watch how long scenes hold together and what Ultra access actually allows once the rollout lands.

Footnotes

1. techcrunch.com, “Google DeepMind Launches Genie 4 World Model in Preview”. Retrieved October 28, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

2. Google DeepMind, “Genie 4: A new generation of world simulation”. Retrieved October 28, 2026. ↩ ↩² ↩³

3. Ars Technica, “DeepMind unveils Genie 4 world model promising more realistic physics and humans”. Retrieved October 28, 2026. ↩ ↩² ↩³

4. The Verge, “Google’s next Genie AI model generates more realistic experiences”. Retrieved October 28, 2026. ↩ ↩² ↩³

5. Google DeepMind, “Genie 3: A new frontier for world models”. Retrieved October 28, 2026. ↩