Genie 4 Lets You Type a World Into Existence — Now It Talks Back, Makes Sound and Obeys Physics

Describe a scene and seconds later you are walking through it, hearing birds, steering a character and arguing with people inside it. Google DeepMind's Genie 4 adds native audio, interactive dialogue and an AI director that keeps the world coherent. Here is what it does, how it works and what it still cannot do.

The AI Spotlight DeskNov 8, 2026 · 11 min read

The stake
What a world model actually is
Everything Genie 4 promises
Build a world on command
Turn a picture into a place
Direct the scene live
Step inside and take control
Characters that talk back
The AI director behind the curtain
Physics that behave
Practice worlds for robots
What Genie 4 still cannot do
Who can use it and when
World model vs game engine vs AI video
Why this matters
How to think about using it
Risks and counterpoints
FAQ
The bottom line

The stake

Type a sentence. A few seconds later, you are standing inside it.

Birds chirp. Light moves across the scene. A character you control walks forward. Someone shouts, "Let's attack the castle!" Someone else warns, "No, it's a trap!" — and the world reacts as you decide what to do.

That is the pitch for Genie 4, the new world model from Google DeepMind, announced on October 27, 2026. DeepMind calls it its most realistic and interactive world model yet.

Genie 3 already rattled the games industry with real-time, explorable AI worlds. Genie 4 adds what was missing: sound, speech, better physics and an AI director that keeps the world from falling apart.

It is still a research preview with clear limits. But it is the clearest look yet at a future where worlds are generated, not built. Here is the full breakdown.

What a world model actually is

A traditional game world is built by hand. Artists model the buildings. Programmers write how doors open and characters behave. Everything you can do was designed in advance.

A world model works differently. You describe what you want, and the model generates the experience frame by frame, in real time, as you explore it. Nothing is pre-rendered. There is no level file. The model predicts what you should see and hear next based on what you are doing.

That is why world models look like video games but are not quite games. They are simulations conjured on demand.

DeepMind frames this as more than entertainment. Understanding and simulating the world, it says, is a crucial part of building artificial general intelligence. A model that can predict what happens next in a realistic world is learning something about how the world works.

Everything Genie 4 promises

DeepMind and coverage from Ars Technica and The Verge describe four headline upgrades:

More realistic worlds. Enhanced photorealism, fluid motion and more natural people.
Real-time changes from natural language. Change the weather, move the camera or add a person mid-simulation.
Native audio. Ambient sound, music, background noise and interactive dialogue, generated with the world.
Gemini reasoning in the loop. A Gemini model plans multi-step events and checks that each step succeeded.

The Verge adds a practical detail: Genie 4 experiences can run about two minutes, up from roughly one minute for Genie 3.

Build a world on command

The simplest use: describe an environment and Genie 4 builds it.

Then you keep going. The model is designed so that you can keep prompting, and Genie 4 works out what happens next. You are not generating a single clip and starting over. You are steering an ongoing world.

You can describe landscapes and events, and the simulation continues from there. DeepMind's framing: 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.

Turn a picture into a place

Genie 4 can also start from an image. Feed it a photo or picture and it transforms it into a dynamic environment ready to explore.

In DeepMind's showcase, a still scene becomes a living one, complete with ambient sound — birdsong rising as the world comes alive.

That is a big deal for creators. A concept painting, a storyboard frame or a location photo can become an explorable space in seconds. DeepMind specifically pitches pre-visualization for filmmakers and prototyping for game ideas.

Direct the scene live

Genie 4 lets you act like a director:

Create a scene.
Move the camera.
Direct every detail.
Watch the world change, live.

The real-time editing works through natural language. Examples DeepMind gives include changing the weather, moving the camera and introducing a new person or object. When you prompt, Gemini contextually adapts your request to fit the surrounding scene, so a new character arrives plausibly rather than popping in at random.

One showcase moment captures the tone: a character asks, "Do you really think we're going to make it?" — and explosions follow.

Step inside and take control

The most game-like mode: describe an interactive experience, and a few seconds later you are inside it with full control over your character.

That is the leap from "AI video" to "AI world." Video models generate something for you to watch. Genie 4 generates something for you to play, responding continuously to your inputs.

The castle scene in DeepMind's showcase is the example. One character urges an attack on the castle; another warns it is a trap and wishes you luck. You decide what happens next.

Characters that talk back

This is the feature DeepMind calls "the true frontier": human speech.

Genie 4 is DeepMind's first real-time world model to generate audio natively. Every world carries ambient sound, and you can specify audio elements in your prompt. On top of that, it introduces interactive dialogue that adapts on the fly. Characters converse and react to your inputs and to events in the world, in real time.

Ars Technica's take: realistic speech was a major shift in realism for regular AI video, and a similar revolution may now be coming to world models — "if it works."

That caveat matters. Realistic, responsive speech in a live simulation is hard. DeepMind lists audio accuracy as an ongoing research challenge.

The AI director behind the curtain

The biggest technical change is how Genie 4 stays coherent.

The model incorporates and builds on the Gemini family in two ways:

Gemini Omni anchors the simulation in real-world appearance and behavior.
Other Gemini models provide the reasoning to keep the world consistent and continuous.

Specifically, Gemini 3.8 Flash acts as a guiding agent in the loop. As Genie 4 generates the world:

Gemini observes the scene continuously.
It breaks down goals into a sequence of steps.
It reviews incoming frames to track progress and verify each step succeeded.
When the world shifts or you intervene, it adapts the plan in the moment.

Between your inputs, Gemini also monitors and steers the scene so it keeps progressing cohesively.

Ars compares this to how reasoning models generate extra internal prompts to land on the right answer. In plain terms: one AI generates the world, another AI watches it and keeps it on script. DeepMind says that is what unlocks reliable cause and effect, procedural behaviors and multi-step events.

Physics that behave

Physics is where AI video often gives itself away: objects float, collide oddly or melt into each other.

DeepMind says Genie 4 is built on an advanced physical understanding of the world and can simulate scenarios involving aerodynamics and collisions. People should look less uncanny, moving naturally and reacting to the changing environment.

Better physics is not just cosmetic. It is what makes a simulation useful for testing anything real.

Practice worlds for robots

That leads to the most consequential use case. DeepMind says Genie 4 can simulate scenarios like helping robots practice tasks virtually, so they might one day carry out that work in the physical world.

The idea: real-world robot training is slow, expensive and sometimes dangerous. A world model can create safe, infinitely variable simulations — training environments for robots and AI agents, and test beds to see how AI systems navigate different situations before they are deployed for real.

DeepMind's showcase ends on a neat line: during all the time spent in other worlds, "our first world kept evolving." The point is that simulated worlds are meant to serve the real one.

What Genie 4 still cannot do

DeepMind publishes a clear list of limitations:

Coherence drifts. Visual details and object permanence can shift over time, especially in longer runs, after many complex edits or while moving around a scene. Ars notes worlds remain consistent for only a couple of minutes.
Multiple agents are hard. Accurately modeling interactions between users, several independent characters and shared environments is an open problem.
Real places are approximate. It cannot simulate real-world locations with perfect accuracy.
Text is weak. Clear, legible text usually appears only if it was in the input description.
Short sessions. It supports a few minutes of continuous interaction, not extended play.
Audio accuracy is still being improved.

Those limits are important. Genie 4 is not a replacement for a game engine. It is a new kind of tool with a very short memory.

Who can use it and when

Right now, not most people.

Genie 4 is rolling out as a research preview for trusted testers.
Next, DeepMind plans to bring it to Google AI Ultra subscribers through Project Genie, and to early-access partners of the Gemini API and Google AI Studio.
There is a waitlist for Project Genie.
Ars notes AI Ultra plans start at $100 per month, developer pricing is not yet published, and access may expand to cheaper plans eventually.

Ars also points out that Genie 3 never went beyond a limited Project Genie web experience, so a broad public release is not guaranteed soon.

DeepMind says Genie 4 includes strict safety guardrails consistent with its other models, went through internal safety evaluations, and was deliberately deployed to trusted testers first.

World model vs game engine vs AI video

It helps to place Genie 4 between two technologies people already know.

A traditional game engine

How worlds are made: by hand, with 3D models, code and level design.
Strengths: perfectly consistent, runs for hundreds of hours, precise rules, readable text and UI.
Weaknesses: slow and expensive to build; you can only do what designers planned.

An AI video generator

How worlds are made: a model renders a clip from a prompt.
Strengths: photorealistic, fast to create, increasingly with sound.
Weaknesses: you watch it; you cannot walk around inside it or change it live.

A world model like Genie 4

How worlds are made: generated frame by frame in real time as you act.
Strengths: interactive, editable with words, sound and speech built in, endless variety.
Weaknesses: minutes of coherence, approximate physics and places, weak text, high compute cost.

The likely future is a blend. Studios could use world models to explore ideas and generate training data, then build the final experience in an engine. Engines may also adopt generative pieces: AI-generated sound, dialogue and background characters inside hand-built worlds.

For robotics, the comparison is different. A physics simulator is precise but narrow; a world model is broad but approximate. Researchers will likely combine both: precise simulators for exact mechanics, world models for the messy variety of the real world — weather, crowds, clutter and surprise.

That is why the "AI director" design matters so much. A generator alone drifts. A generator paired with a planner that checks its work is the first step toward simulations that can be trusted for longer than a couple of minutes.

Why this matters

For games: Genie 3's reveal reportedly sent gaming stocks tumbling. Genie 4 makes the threat — or opportunity — more concrete. Prototyping a game idea by describing it, then playing it seconds later, changes early-stage design. It does not replace crafted games yet, given the two-minute coherence limit.

For film and media: pre-visualization becomes conversational. Directors can scout a location that does not exist.

For robotics and AI agents: realistic, interactive, variable simulations are exactly what embodied AI needs to train safely.

For AI research: pairing a generator with a reasoning model that plans and verifies is a powerful pattern. Expect it in many other systems.

For AGI debates: if predicting the world is a core part of intelligence, world models are one of the most direct tests of how much these systems understand.

How to think about using it

If you get access, treat Genie 4 as a sketchbook, not a studio:

Prototype, don't ship. Use it to test ideas, moods and mechanics quickly.
Start from images you already have: concept art, storyboards, photos.
Keep sessions short and focused on one idea.
Iterate with prompts — weather, camera, new characters — rather than starting over.
Capture what works and rebuild it properly in a traditional pipeline.
Do not rely on text or real-world accuracy inside the simulation.

Risks and counterpoints

Demos are curated. Showcase clips are the best outcomes, not the average.
Short memory. A couple of minutes of coherence limits real use.
Cost and access. Premium pricing and phased rollout mean most people will not touch it soon.
Compute and energy. Real-time photorealistic generation with an AI director in the loop is expensive to run.
Misuse potential. Realistic people and speech raise deepfake concerns; DeepMind says guardrails are in place, but they are untested at scale.
Creative labor. If worlds can be generated, questions about artists' jobs and training data will intensify.

FAQ

What is Genie 4?

A real-time world model from Google DeepMind that generates interactive, explorable environments from text or images, with native audio, dialogue and physics.

How is it different from Genie 3?

More photorealism, better physics, native audio and speech, Gemini-guided coherence, and experiences of about two minutes instead of roughly one.

Can I play Genie 4 now?

Only trusted testers can for now. DeepMind plans to expand access to Google AI Ultra subscribers through Project Genie and to API partners. There is a waitlist.

Will it replace video games?

Not soon. Coherence lasts only minutes, multi-character interactions are hard, and text rendering is weak. It is best for prototyping and research.

Why does DeepMind care about world models?

It sees simulating the world as key to AGI, and as a way to create safe training environments for robots and AI agents.

The bottom line

Genie 4 turns a sentence or a picture into a world you can walk through, hear and talk to — and a second AI keeps that world on track as you poke at it.

It is not a game engine. It forgets itself within minutes, struggles with crowds and text, and most people cannot use it yet. But the direction is unmistakable. Worlds used to be built. Now they are being generated, live, with sound and speech. What happens next, as DeepMind puts it, is up to you.