Genie 4 Explained: The World Model You Can Talk To, Steer Live, and Hear
Type a sentence. Get a world you can walk around in. Type another sentence and change the weather while you are standing in it — while the people in that world keep talking.
That is the pitch behind Genie 4, Google DeepMind’s new world model, announced on 27 October 2026. It is not a video generator with a new name, and it is not a game you can download today. It is a research preview, available only to a small group of trusted testers, and it is the clearest sign yet of where interactive AI is heading: from clips you watch to worlds you step into.
First, what is a world model?
A video model produces a clip, and you watch it. A world model produces a place, and you move through it. Press a key, the camera moves. Type a change, the world changes. Behind the scenes it is still generating frame by frame — but every frame responds to what you just did, instead of following a script written in advance.
That distinction is the whole story of Genie 4. DeepMind’s earlier Genie models already did the visuals. Genie 4 adds the three things that make a generated place feel like a place: sound, people who behave, and the ability to change it while it runs.
The four advances, in DeepMind’s own framing
DeepMind’s announcement post sets out four capabilities. Here they are, in plain language:
- You can reshape the world while it runs. With natural-language prompts you can change the weather, move the camera, or introduce a new person or object mid-simulation — and the scene keeps going instead of restarting. Gemini adapts your prompt to fit the surrounding scene, and between your inputs it monitors and steers the scene so the world keeps progressing cohesively. Navigation uses standard keyboard controls.
- Physics and people take a visible step up. Genie 4 simulates real-world physics scenarios such as aerodynamics and collisions. But DeepMind calls the simulated people “the most dramatic leap”: characters that move naturally, respond organically to their environments, and can be interactively steered in real time.
- The world now has sound — including speech. Genie 4 is DeepMind’s first real-time world model to generate audio natively. Every world carries the ambient sound of its own setting, from wind in the trees to waves crashing at a beach. The bigger frontier, in DeepMind’s words, is human speech: characters hold conversations and react to your inputs or to events in the world, all in real time. The silence that made earlier Genie clips feel eerie is gone — at least in DeepMind’s demos.
- Gemini now reasons inside the simulation. This is the architectural change that matters most for anyone building with AI. Gemini 3.8 Flash serves as a guiding agent in the loop: as Genie 4 generates the world, Gemini continuously observes the scene, breaks a goal down into a sequence of steps, reviews incoming frames to track progress, and verifies that each step succeeded. When the world shifts or you intervene, Gemini adapts its plan in the moment. Set an event in motion and Genie 4 simulates what happens next — how the world responds, how it sounds, and how the people in it move, react, and speak.
The robot experiment that shows why this matters
DeepMind pairs Genie 4 with another frontier model, Gemini 4 Argon, in an experiment tied to Gemini Robotics. Argon breaks a user-instructed task into steps, and Genie 4 renders each step in real time — in the published example, Apptronik’s Apollo humanoid playing a game of checkers.
Why show checkers? Because it is a stand-in for something much bigger: testing how an AI system navigates a task in simulation before anything happens in the real world. If a robot can rehearse a job inside a world that looks, sounds, and behaves like the real one, the expensive and dangerous part of training moves into simulation. That is the same logic behind using world models to train agents — and it is why DeepMind frames world simulation as “a crucial aspect of building AGI”, not just as entertainment technology.
What trusted testers are actually doing with it
DeepMind has been working with tester cohorts across creative industries, journalism, education, gaming, and engineering, and their reported uses are more concrete than the usual demo hype:
- Filmmakers use it as a rapid “draft mode” for pre-visualisation: step into a live simulation of a scene and run multiple “takes” in real time, adjusting camera angles, lighting, dialogue, and action on the fly.
- Photojournalists are testing how a single 2D archival photograph can become a grounded simulation of that environment in seconds — and how populated, dynamic worlds can help audiences experience historical spaces, or visualise how a short-term event like a flash flood or wildfire unfolds.
- Game developers go from an initial idea to an explorable prototype in seconds, testing visual styles, material aesthetics, and interactive mechanics through natural-language prompts before writing a single line of engine code.
- Enterprise partners are evaluating it for rapid “zero-to-one” scenario generation — sketching a situation that does not exist yet, then probing it.
None of these are finished production pipelines. They are the earliest workflows, from people DeepMind chose to test with first.
Genie 4 vs Genie 3: what actually changed
| Genie 3 (Aug 2025, reported) | Genie 4 (27 Oct 2026) | |
|---|---|---|
| Sound | None — silent clips | Native audio: ambient sound, music, interactive dialogue, generated with the world |
| Steering | Prompt at the start; limited mid-run changes | Natural-language edits mid-simulation, Gemini adapts them to the scene |
| People | Basic characters | Characters that move naturally, respond organically, can be steered live |
| Physics | Basic (e.g. gravity) | Aerodynamics, collisions, more realistic interaction |
| Scene length | ~60 seconds (reported) | ~2 minutes (TechCrunch-reported, not a DeepMind spec) |
| Guiding intelligence | Agent navigating inside the sim (SIMA 2, reported) | Gemini 3.8 Flash guiding the generation itself, in the loop |
Two honesty notes on that table. First, the “two minutes” figure comes from TechCrunch’s briefing, not from DeepMind’s blog — DeepMind itself publishes no duration number. Second, Genie 3’s specs (720p, 24 frames per second, silent) are as previously reported, not re-verified here. The direction of travel is unambiguous even with those caveats: longer, louder, more steerable, more human.
Can you try it? Almost certainly not yet — and that is deliberate
Genie 4 is a research preview for trusted testers. DeepMind says it is coming soon to Google AI Ultra subscribers in Project Genie, and to early-access partners of the Gemini API and Google AI Studio. A waitlist is open through Project Genie. There is no public launch date and no announced price for Genie 4 itself — so if any roundup quotes you one, it is guessing.
DeepMind is explicit about why the rollout is staged: Genie 4 merges real-time generation with photorealism, which it calls “a new area of frontier research”, and it wants feedback from trusted partners before a broader release. The model went through internal safety evaluations first, and it carries strict guardrails consistent with DeepMind’s other models.
The limits DeepMind admits itself
This is the part that separates a serious announcement from a hype reel, and DeepMind states it plainly:
- Visual details and object permanence can still shift or drift over time — especially as a simulation runs longer, after multiple complex prompt edits, or while navigating through the scene.
- Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are all active areas of ongoing research.
- It maintains coherent environments over short sessions. Two minutes of coherence is a scene, not a game session and not a film.
In other words: the demos are real, but they are short, curated, and run under DeepMind’s supervision. Anyone telling you Genie 4 is a finished game engine or a production film tool is ahead of the evidence.
The safety line worth knowing
To maintain transparency, DeepMind says all video and audio generated by Genie 4 will include SynthID, its imperceptible digital watermarking technology for identifying AI-generated content across visual and audio modalities. As generated worlds get more photorealistic — and now audible — that watermark is doing more work than ever. It is also a reminder: the most convincing demo clip you see from Genie 4 this week is still, by design, traceable as synthetic.
Why this matters beyond games
The crowded field tells its own story. In the same season as Genie 4, Runway released its second world model (GWM Worlds 2, also playable and with audio), World Labs is being acquired by AMD, Amazon is reported to be in talks to buy Decart, and Rivian acquired Third Dimension for in-house simulation. Everyone is converging on the same bet: the next platform is not a better clip, but a simulatable world.
DeepMind’s longer-term pitch is “hybrid world models”: combining the logical, structured reasoning of code and language models with the rich, sensory representation of pixel-level simulation. Gemini maps the structure and plans the narrative timeline; Genie renders the immediate physical reality, pixel by pixel. If that hybrid works, the result is a world model far more coherent, controllable, and grounded than either approach alone — and that is a tool for training robots, testing agents, prototyping experiences, and eventually, building interactive media that is generated rather than authored.
The honest verdict
Genie 4 is a genuine milestone with clearly labelled limits: the first DeepMind world model with native sound and dialogue, live steering, and Gemini reasoning inside the loop — wrapped in a staged, testers-only preview with admitted drift, no public date, and no price. Watch the official DeepMind video below, and the Project Genie tutorial after it, and judge it when you can actually step inside a world yourself. Until then, treat every viral clip as a two-minute scene, not a finished product.
Official Google DeepMind video: “Genie 4: A new generation of world simulation” — shows real-time Genie 4 interactions, keyboard navigation, and text-prompt steering.
Project Genie tutorial: “Project Genie Tutorial (How to use Project Genie)” by TheAIGRID — a walkthrough of Project Genie, the app Genie 4 is rolling out to. Honest note: Genie 4 itself is not publicly usable yet, so no hands-on Genie 4 tutorial exists; this tutorial covers the app it will arrive in, including access requirements, world creation, and exploration.
FAQ
What is Genie 4?
A world model from Google DeepMind that generates interactive, explorable worlds in real time from a text or image prompt — with native audio, steerable characters, and live text steering — announced 27 October 2026.
Is Genie 4 available to the public?
No. It is a research preview for trusted testers. It is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. A waitlist is open.
How is a world model different from a video generator?
A video generator produces a clip you watch. A world model produces a place you move through: every frame responds to your keyboard input and prompt edits as they arrive.
How long can a Genie 4 scene run?
TechCrunch reports coherent scenes of up to about two minutes, roughly double Genie 3. DeepMind’s own blog gives no duration figure, and it warns that details and object permanence can drift over longer runs.
Does Genie 4 generate sound?
Yes — it is DeepMind’s first real-time world model with native audio: ambient sound, music, and interactive dialogue generated together with the visuals. All Genie 4 video and audio carries SynthID watermarking.
Can Genie 4 be used for robotics training?
That is one of DeepMind’s stated directions: in an experiment with Gemini 4 Argon, Genie 4 simulated a humanoid robot’s task step by step in real time, as a path to rehearsing scenarios in simulation before the real world.
Primary source: Google DeepMind, “Genie 4: A new generation of world simulation” (27 Oct 2026). The two-minute scene figure is TechCrunch-reported, not a DeepMind specification. Images are from Google DeepMind’s Genie 4 announcement page. The embedded videos are Google DeepMind’s official Genie 4 video and TheAIGRID’s Project Genie tutorial. The idea for this article came from a Unite.ai news item; no text was copied from it.
