Genie 4 builds striking worlds, but the AGI claim needs more than a demo

Google DeepMind put out Genie 4 on Tuesday, October 27, as a research preview. The model generates interactive worlds you can walk through, with more photorealistic scenes, better physics and more humanlike characters than its predecessor. ¹²

The new features are real. Genie 4 generates audio, including music, ambient noise and spoken dialogue that adapts as the scene changes. ¹² Users can type prompts mid-simulation to alter the weather, move the camera or add people and objects. ¹ Under the hood, Gemini 3.8 Flash acts as a guide that breaks goals into steps and checks each frame. ³²

Now the limits. A scene holds together for about two minutes, double what Genie 3 managed. ¹³ Past that, details and object permanence shift, objects drift, and the team says it has not worked out how to make audio reliably accurate or in-world text legible. ¹³ I have run systems that were declared production-ready after a two-minute test.

Google frames world models as a key stepping stone toward artificial general intelligence, because they could supply unlimited training environments for AI agents. ⁴² That is a coherent research argument. It is also a claim about the future, and the present product is a simulator that forgets how a scene began once the clock passes two minutes. A convincing fake world is a rendering achievement. Whether it means the model understands the real one is exactly what the drift problem calls into question.

There is a narrower, more believable payoff. Google is pitching pre-visualization for filmmakers, game prototyping and robot training scenarios. ³ In one experiment tied to Gemini Robotics, a Gemini model split tasks into steps that Genie 4 simulated in real time, a possible way to test robot scenarios before they happen in the real world. ¹ Trusted testers also point to storyboarding and rapid ideation, where staying in a creative flow matters more than long-term consistency. ¹

For games, Genie 3 drew notice last year because its output looked like a video game. ² Genie 4 is more convincing, but a two-minute window with drifting objects does not yet replace level design, saved state or rules that stay put. Prototyping is the realistic near-term use.

Access is limited. Genie 4 is available now to trusted testers, with a rollout coming to Google AI Ultra subscribers in Project Genie and to early partners of the Gemini API and Google AI Studio. ¹ Ultra plans start at $100 per month, and pricing for developers has not been announced. ² A waitlist is open. ³

The competition is moving too. Runway released a second world model with audio a couple of months ago, and the field is consolidating through acquisitions and deals. ¹

Our read: Genie 4 is a real step for simulation, and the audio and live steering are the parts worth watching. The AGI label is Google’s framing, not something the demo establishes. Judge it by how long a world stays coherent, and by whether the drift and audio problems get fixed in the next release.

Footnotes

1. TechCrunch, “Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

2. Ars Technica, “DeepMind unveils Genie 4 world model promising more realistic physics and humans”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. The Verge, “Google’s next Genie AI model generates more realistic experiences”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Google DeepMind, “Genie 3: A new frontier for world models”. Retrieved October 31, 2026. ↩