What Is Genie 4? Google DeepMind’s World Model Adds Sound, Dialogue and Live Edits

Genie 4 is Google DeepMind’s newest world model, announced on October 27, 2026. It generates an interactive world from a text description, and it now adds sound and spoken dialogue. You can also change the world while you are in it by typing what you want.

Most of our readers will not get to use it yet. Genie 4 is a research preview for trusted testers. Google says wider access is "coming soon," starting with its top consumer plan and some developer partners. This post covers what Genie 4 is, what changed since Genie 3, who can use it, what DeepMind says it is for, the limits DeepMind lists, and where it might fit for agencies and nonprofits.

The short version: Genie 4 is a real step for world models. Sound and dialogue now come from the same model as the picture. Its sessions are also short. DeepMind says it holds together for a few minutes, drifts over longer sessions, and cannot yet recreate real places accurately. Watch it and test it, but do not plan client work around it yet.

What Genie 4 is

A world model is an AI system that simulates an environment you can move through and act in. DeepMind’s Genie 3 announcement described world models as "AI systems that can use their understanding of the world to simulate aspects of it."

A video generator gives you a finished clip. A world model keeps generating as you act. You walk forward, and it draws what is ahead. You turn around, and it has to remember what was behind you.

According to DeepMind’s Genie 4 model page, it is the lab’s "most realistic and interactive world model yet." The page says you can create and change landscapes, events, camera movement, characters and weather with plain language. It describes worlds where objects collide, water flows and characters move naturally.

If the difference between this and a chatbot is new to you, our guide to large language models covers the text side.

What is new in Genie 4 compared with Genie 3

DeepMind’s posts do not include a side-by-side comparison. The table below lines up what each announcement says.

Three changes stand out.

Native audio. DeepMind’s Genie 4 announcement says, "Genie 4 is our first real-time world model to generate audio natively." The model page gives examples: leaves crunching, crowds cheering, characters speaking.

Dialogue. The announcement says, "Genie 4 introduces interactive dialogue that adapts on the fly." Characters can talk, and what they say can respond to what happens.

Live editing. Genie 3 already had promptable events. Genie 4’s announcement says "you can use natural language to modify a simulation in real-time." Its examples include "changing the weather, moving the camera, or introducing a new person or object."

How Gemini fits into Genie 4

Genie 4 builds on Google’s Gemini models. The announcement says "Gemini Omni anchors the simulations in real-world appearance and behavior." Other Gemini models handle coherence and continuity.

It also says "Gemini 3.8 Flash serves as a guiding agent in the loop." In that role, Gemini breaks a goal into steps and checks progress inside the world.

The model page adds a robotics example in which Gemini 4 Argon writes a multi-step plan for a task. We covered those models in our posts on Gemini Omni Flash, Gemini 4 Argon and Gemini 3.8 Flash costs.

Who can use Genie 4 right now

As of early November 2026, access is narrow. DeepMind says Genie 4 is rolling out "as a research preview for trusted testers."

The announcement names two next steps, both "coming soon":

• Google AI Ultra subscribers, through Project Genie.
• Early access partners of the Gemini API and Google AI Studio.

A waitlist is open through the model page. DeepMind has not published a date, an API price or a model ID for Genie 4.

It helps to know what Project Genie is today. Google’s January 29, 2026 Project Genie post calls it an "experimental research prototype." It ran on Genie 3 and opened to Google AI Ultra subscribers in the US who are 18 or older. That post says generations are capped at 60 seconds, and Google AI Ultra for Business accounts were left out.

Google’s subscriptions page still describes Project Genie as running on Genie 3, as of early November 2026. It lists Google AI Ultra from $99.99 a month, with a higher tier at $199.99. If Genie 4 replaces Genie 3 there, these details may change. For developers, our Google AI Studio 101 shows where an AI Studio release would appear.

What DeepMind says Genie 4 is for

The announcement lists several uses. Preview partners are testing some of them. These are DeepMind’s own descriptions.

• Pre-visualization. DeepMind says “Creatives and technologists are testing Genie 4 as a rapid ‘draft mode’ for pre-visualization and ideation.”
• History from photos. It says “a single 2D archival photograph can become a grounded simulation of that environment.”
• Events like floods and fires. It describes using Genie 4 to “visualize how short-term environmental events, like a flash flood or wildfire, unfold.”
• Game concepts. Designers can “go from an initial idea to an explorable prototype in seconds.”
• Enterprise scenarios. Partners are evaluating it for “zero-to-one” scenario generation.
• Robotics research. One demo shows Gemini 4 Argon and Genie 4 simulating Apptronik’s Apollo humanoid playing checkers.

The model page also points to training agents such as SIMA, DeepMind’s agent for 3D virtual worlds. That is the long-term goal. The page says exploring many realistic environments is "a key stepping stone on the path to AGI." Our explainer on AI agents covers the software side of that idea.

The limits DeepMind lists for Genie 4

DeepMind is direct about what the model cannot do yet. The model page lists these limits:

• Drift. “Visual details and object permanence can shift or drift over time,” especially in longer sessions or after many edits.
• Session length. It supports “a few minutes of continuous interaction, rather than extended sessions.”
• Real places. It cannot yet simulate real locations with perfect accuracy.
• Text. Legible text “is often only generated when it’s in the input world description.”
• Multiple agents. Complex interactions between users and several agents remain “an ongoing research challenge.”
• Audio. “Real-time audio accuracy remains an area of ongoing research and improvement.”

The announcement adds complex physical interactions and spatial accuracy to that list. It calls Genie 4 "an early, evolving area of research."

On safety, DeepMind says Genie 4 has guardrails, went through internal safety reviews, and went to trusted testers first. It also says "all video and audio generated by Genie 4 will include SynthID," Google’s invisible watermark.

The physics and realism claims come from DeepMind’s own demos. As we noted in our post on vendor benchmarks, treat a vendor’s own showcase as a starting point until others test it.

Where Genie 4 might fit for agencies and nonprofits

None of the uses below are proven yet for our readers. They are places to test Genie 4 once you can get in.

Previsualization for campaigns and events. An agency could sketch the feel of a video shoot or an exhibit space before anyone books a crew. A rough, explorable draft can settle questions about mood and layout that a mood board leaves open. The few-minute session limit fits this kind of quick look.

Prototypes for interactive learning. Museums, libraries and education nonprofits often pitch interactive experiences to funders. A short walkthrough could show the idea faster than a written concept. It would be a pitch aid. The drift and text limits mean it is no substitute for a built experience.

Training and awareness simulations. DeepMind names flash floods and wildfires. A disaster-readiness group could use a short simulation to show how a scene changes. DeepMind’s own limits matter most here. The model cannot reproduce real places accurately, and its physics are unverified. Do not present a generated world as a model of what will actually happen at a real site.

Two practical points apply to all three. SynthID marks the output as AI-generated, so plan to disclose it the way you would any AI media. And until Google publishes terms and pricing, you cannot quote it in a proposal.

How Genie 4 compares with other world-model work

Other labs are working on the same problem from different angles. OpenAI framed Sora 2 in September 2025 as a step toward "general-purpose world simulators," with synchronized dialogue and sound. That page now says the Sora product has not been available since April 26, 2026. Meta’s V-JEPA 2, released in June 2025, is a world model built for robot planning. It does not produce explorable scenes. NVIDIA’s Cosmos offers open world foundation models for physical AI. World Labs’ Marble builds full 3D worlds you can export as splats or meshes. DeepMind does not describe an export like that for Genie 4. Our physical AI devices landscape covers the hardware side of the same trend.

What to do about Genie 4 now

1. Join the waitlist on DeepMind’s model page if you want early access.
2. Watch the Gemini API and AI Studio for an early access program, model ID and terms.
3. Pick one test case, such as an event walkthrough or exhibit pitch, so a trial has a clear goal.
4. Keep claims modest with clients until you have run it yourself.

Frequently Asked Questions