What is Genie 4? Google's world model with audio

What is Genie 4?

DeepMind's new world model generates an explorable world, along with ambient sound and character dialogue, in real time as you move and type. Genie 4 comes from Google DeepMind, Google's AI research lab. DeepMind says it is the lab's first real-time world model to generate audio natively.

Under the hood, Genie 4 relies on Google's Gemini models. Gemini Omni keeps appearance and behavior close to the real world, and other Gemini models handle the reasoning that keeps a simulation coherent from moment to moment.

A world model draws the next frame from your input

Instead of rendering a fixed clip like a video generator, it keeps predicting what you would see after each key press or text prompt. Walk forward and the road ahead appears. Turn around and the view behind you is generated on the spot.

In Genie 4, the keyboard moves you and the camera, and text prompts change what is in the scene. When you trigger an event, the model predicts how the world reacts, what it sounds like and how people in it move and speak.

DeepMind released it as a research preview on October 27

Only trusted testers can use it so far, and DeepMind has given no date for a public release. A research preview means a small group tries the model before it becomes a product. Genie 4 follows Genie 3, which DeepMind announced on August 5, 2025.

What's new: text edits, physics, audio and Gemini planning

DeepMind lists four advances over earlier versions, and a Gemini model takes part in every one of them. Genie 4 draws the world, while Gemini interprets your prompts and tracks progress. Of the four, native audio is the one Genie 3 did not have at all.

Four advances in Genie 4

Advance	What it does
Text edits	Change the weather, camera, people or objects with a prompt
Physics and people	Aerodynamics, collisions and characters that react to their surroundings
Native audio	Ambient sound for each place and dialogue from characters
Complex events	Gemini 3.8 Flash splits a goal into steps and checks each one

Change the weather, camera or cast with a text prompt

Gemini first adjusts your request so it fits the surrounding scene, then the world updates without being rebuilt. You can type a prompt while you are still walking around, and the change appears in the same session.

Between prompts, Gemini keeps watching the scene and steering it so the story does not fall apart. That is why a new character added mid-session behaves as if it belongs there.

Collisions, aerodynamics and more lifelike people

DeepMind calls characters its biggest improvement: they move naturally, react to what happens around them and can be controlled live. On the physics side, Genie 4 simulates aerodynamics and collisions, so objects fly and bump into each other the way you would expect.

Interactions between several characters were still an open research problem when DeepMind released Genie 3. Genie 4 is the first version where people in the world talk and react to each other as part of the simulation.

Ambient sound and dialogue are generated with the world

Each place gets its own soundscape, like wind in the trees or waves on a beach, and characters answer your input as it happens. The audio is generated together with the video rather than added afterward.

Characters respond to what you type and to events in the world. Because they are not reading from a script, the same world can produce a different conversation each time you change what you do.

Gemini 3.8 Flash breaks complex events into steps

Gemini 3.8 Flash watches the scene, splits a goal into a sequence of steps and drives the simulation through them. DeepMind paired Genie 4 with Gemini 4 Argon to simulate Apptronik's humanoid robot Apollo playing a game of checkers.

In DeepMind's demo, the guiding model checks each incoming frame to confirm a step worked before moving Apptronik's Apollo robot to the next checkers move. That check keeps a long, multi-step event from going off track.

How Genie 4 compares with Genie 3

The new model adds native audio and characters that talk and react, while text-driven world events already existed in the previous version. DeepMind called that Genie 3 feature "promptable world events," and it let you change the weather or add objects with a sentence.

Published figures for Genie 3 and Genie 4

Item	Genie 3	Genie 4
Announced	August 5, 2025	October 27, 2026
Resolution	720p	Not published
Frame rate	24 fps	Not published
Consistent for	A few minutes	Not published
Audio	None	Music, ambient sound and dialogue
Text edits	Yes	Yes, adjusted by Gemini
Multi-character interaction	Open research problem	Characters talk and react

Only Genie 3 has published numbers. As of October 31, 2026, DeepMind has not released matching figures for Genie 4, so the table cannot say which model is sharper or runs longer.

Genie 3 ran at 720p and 24 fps with no sound

Genie 3 generated worlds at 720p, the resolution streaming services label HD, at 24 frames per second. It was DeepMind's first world model you could control in real time.

The previous model kept a world consistent for only a few minutes, and DeepMind listed interactions between multiple characters as unsolved. It also produced no audio, so every world was silent.

DeepMind has not published resolution or session length

Genie 4's announcement gives no resolution, frame rate, session length or price. Some outlets report that Genie 4 holds a scene for longer than Genie 3, but that figure does not appear in DeepMind's post.

The only official number to compare against is Project Genie's 60-second limit for exploring a world, which applies to Genie 3 today. Whether Genie 4 keeps the same limit will be clear once it reaches Project Genie.

When can you use Genie 4?

Access widens in three stages, from trusted testers to Ultra subscribers and early access developers, and then to everyone. DeepMind has not put a date on any of them.

Genie 4 rollout order

Stage	Who	Status (October 31, 2026)
1	Trusted testers	Research preview, live
2	Google AI Ultra subscribers in Project Genie	Coming soon, waitlist open
2	Early access partners in the Gemini API and Google AI Studio	Coming soon
3	Everyone else, users and developers	Later

Only trusted testers have access today

Those testers work in five fields: creative industries, journalism, education, gaming and engineering. DeepMind has not explained how testers are picked. For everyone else, the one thing to do now is join the Project Genie waitlist.

Google AI Ultra subscribers get it next in Project Genie

You can already join a waitlist, and the help page limits Project Genie to adults 18 or older in countries where the plan is sold. Project Genie is Google's research prototype for building worlds from text or images, and it runs Genie 3 today. Google AI Ultra is the company's top paid AI subscription.

According to Google's help page, generating a world in Project Genie does not use AI credits. Each world gives you 60 seconds of exploration.

Developers get early access through the Gemini API

The developer route runs through the Gemini API, which lets apps call Gemini models from code, and Google AI Studio, the browser tool for trying those models. Early access partners there get Genie 4 alongside Ultra subscribers.

DeepMind has not said how partners are chosen or what usage will cost, and wider developer availability comes after that phase. Until then, there is no public API price to plan around.

What testers use it for and where it falls short

Every published use case is a quick draft before real production, which fits a model whose worlds drift over long sessions. DeepMind describes those limits itself in the announcement.

Previsualization, old photos and game prototypes

Testers use Genie 4 to plan film shoots, rebuild places from old photos and prototype games. DeepMind gives four examples.

How testers use Genie 4

Tester	Use
Filmmakers	Previsualization, changing camera position, lighting, dialogue and movement between takes
A photojournalist	Turning one old photo into a place you can walk through within seconds
Game developers	Turning an idea into a walkable prototype before writing engine code
Enterprise partners	Evaluating it for simulations built from scratch

In each case the goal is to see something close to the finished look in seconds, before filming or coding starts. None of DeepMind's examples use Genie 4 to make the final film or game.

Details drift during long sessions

Short sessions hold together, but long play, stacked prompts or wide exploration can make objects shift or vanish. Something you passed earlier may look different, or be gone, when you come back to it.

DeepMind has not said when these problems will be fixed. It lists five areas it is still working on:

• Complex physical interactions
• Audio accuracy
• Legible text
• Spatial accuracy
• Consistency over long sessions

In practice, do not count on readable signs or long working sessions inside one world with this version. Both appear on that list, as legible text and long-session consistency.

SynthID watermarks every video and audio output

Every video and audio clip Genie 4 generates carries SynthID, Google's watermark for AI-made content. DeepMind also ran internal safety evaluations and added safeguards before the release.

The mark is invisible to people but lets detection tools confirm later that the content came from AI. Google's SynthID Detector is one such tool, covered in "SynthID Detector: How to Check for AI Watermarks."

Join the Project Genie waitlist and wait for the numbers

The model now sounds and talks, but its specs, price and public release are still unannounced. The tester use cases and the drift problem point the same way: Genie 4 suits short previews, not hours of work inside one world.

If you already pay for Google AI Ultra, the waitlist is the fastest way to see how far Genie 4 goes beyond Project Genie's current 60 seconds. If you do not, nothing published so far gives a reason to subscribe for Genie 4 alone before DeepMind shares its numbers.