Genie 4 puts an agent inside a world model

Google DeepMind announced Genie 4 on Tuesday 27 October 2026 and called it "our most realistic and interactive world model" ^[1]. You describe a place in plain words and it builds a scene you can move through, frame by frame and in real time, and you can keep typing while it runs to change the weather, move the camera or add a person or an object ^[1]. It's a research preview for trusted testers, so I haven't used it, and everything below is from Google's post and the coverage of it.

What's new since Genie 3?

Genie 3 came out as a research preview in August 2025 ^[1]. Genie 4 keeps a scene consistent for up to two minutes, twice as long ^[2]. It's the first of Google's real-time world models to make sound itself, music, background noise and characters who hold a conversation that changes as you talk to them ^[1]. Google says the physics and the way people move are better, with aerodynamics and collisions in the demos, and that Gemini Omni keeps the scenes looking and behaving like the real world ^[1].

The change I find most interesting is that Gemini now runs inside the loop. Google's post says Gemini 3.8 Flash watches the scene, breaks a goal into steps, checks each new frame to see whether a step worked, and changes the plan when the world shifts or someone steps in ^[1]. That's an agent (observe, plan, act, check), only the world it acts in is being generated around it.

Why does a world model matter for agents?

Agents learn from data today, and most of it is text. A world model is somewhere an agent can try things and see what happens, and The Decoder's headline calls Genie 4 a modern spin on agent designs that predate language models ^[3]. The Decoder put it as world models becoming "training environments where robots and AI agents learn from simulated experience" ^[3]. Google showed a version of that, with Genie 4 simulating Apptronik's Apollo humanoid playing checkers while a Gemini Robotics model gave the steps, and no real robot in the room ^{[1] [3]}.

I think this is where world models end up mattering most for what we track here, as a test track where an agent can fail cheaply before it touches anything real. Two minutes of consistency is short for that, and Google says the details drift in long sessions.

What doesn't work yet?

Google lists it plainly. Details and objects drift the longer a session runs or the more you edit it, and complex physical interactions, the accuracy of the sound, readable text inside the world, spatial accuracy and long-term consistency are all still research problems ^[1]. The team told TechCrunch it's early and that the end product isn't settled yet ^[2]. Everything it makes carries a SynthID watermark, in the picture and the sound ^[1].

Who can use it?

Trusted testers have it as a research preview, and Google says it's coming to Google AI Ultra subscribers in Project Genie, which has a waitlist, and to early access partners of the Gemini API and Google AI Studio ^[1]. TechCrunch reports the rollout as over the coming weeks ^[2]. There's no price beyond the Ultra subscription.

A couple of months earlier Runway released GWM Worlds 2, a playable world model with sound, and TechCrunch's piece lists World Labs being bought by AMD, Amazon reportedly in talks to buy Decart, and Rivian buying Third Dimension for self-driving simulation ^[2].

Genie 4 is on our Coming models page. We list things when someone outside the company can call them, so it gets a listing when the Gemini API access opens, and I'm curious to see what it costs per minute of world.

The terms, explained

References

Read on 27 October 2026. Quotes are as published.

1. [1]Google DeepMind, "Genie 4, a new generation of world simulation", 27 October 2026. Source for what Genie 4 does, real-time steering, the sound and dialogue, physics and human motion, Gemini Omni and Gemini 3.8 Flash's planning role, the Apollo checkers demo, the research preview, Project Genie and its waitlist, the Gemini API and AI Studio, SynthID, the stated limitations and Genie 3's August 2025 preview. deepmind.google
2. [2]Rebecca Bellan, "Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering", TechCrunch, 27 October 2026. Source for two minutes of consistency against Genie 3, the rollout over the coming weeks, the team calling it early, and Runway, World Labs, Decart and Third Dimension. techcrunch.com
3. [3]Maximilian Schreiner, "Google Deepmind's Genie 4 puts a modern spin on agent designs that predate LLMs", The Decoder, 27 October 2026. Source for world models as "training environments where robots and AI agents learn from simulated experience" and the robot demo's details. the-decoder.com