Google DeepMind unveils Genie 4: real-time worlds with sound and dialogue, steered by text

Google DeepMind has launched Genie 4, a new world model that generates photorealistic, interactive worlds in real time. Users can change the weather, move the camera or add new people with plain text while the simulation runs, and the model generates sound, music and dialogue alongside the picture. For now, Genie 4 is open only to a limited group of testers.

Google DeepMind posted the announcement on X at 17:10 Oslo time (16:10 UTC) on Tuesday 27 October, alongside a blog post the same day. Google CEO Sundar Pichai called it the company's most realistic and interactive world model yet.

Steered with text while the world runs

A world model simulates an environment frame by frame and reacts to what the user does. In Genie 4, users navigate with keyboard and camera controls and steer the action with text prompts as it unfolds. According to DeepMind, Gemini adapts each prompt to the scene, and between prompts Gemini monitors the simulation to keep the world coherent.

The model builds on the Gemini family. DeepMind writes that Gemini Omni anchors the simulation in how the world actually looks and behaves, while Gemini 3.8 Flash acts as a guiding agent: it breaks goals into steps, watches incoming frames and checks whether each step succeeded.

Sound and people are the big changes

DeepMind says Genie 4 is its first real-time world model to generate audio natively. Each world carries the sound of its own setting, such as wind in the trees or waves on a beach. Characters in the simulation can also talk and respond to what the user does, in real time.

DeepMind describes the way the model simulates people as the biggest leap: characters are meant to move naturally and react to their surroundings. The company also reports improved physics, such as aerodynamics and collisions.

In one research example, DeepMind paired Genie 4 with Gemini 4 Argon. Argon broke a task into steps, and Genie 4 simulated Apptronik's Apollo humanoid robot playing checkers. Google presents this as a way to test robots and agents in simulated scenarios before they are released into the physical world.

Limited access and known weaknesses

Genie 4 has been rolled out as a research preview to selected testers. Google says the model is coming "soon" to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. A waitlist is open. The announcement does not state pricing, a date or which countries will get access.

DeepMind is open about the limitations. Details and objects can shift or disappear as a simulation runs longer, after several complex prompt edits or while the user moves around the scene. Physics, audio, legible text and spatial accuracy over time remain research areas. All video and audio from Genie 4 is marked with Google's imperceptible SynthID watermark, and the company says the model follows the same safety policies as its other models.

What it means for leaders

According to DeepMind, testers are using Genie 4 for film pre-visualisation, for turning archival photos into simulated environments, for game prototypes and for building training and simulation scenarios in companies. For businesses in media, marketing, gaming, training and robotics, tools like this could shorten the path from an idea to something people can see and test.

It is still a preview with no price or date. Teams that want to try it can join the waitlist, but should hold off on plans that assume the tool is available. Photorealistic, steerable worlds with natural-looking people and voices also make it more important to have rules for labelling AI-generated content and for what employees may create in the company's name.

Sources and media

• Google DeepMind: Genie 4: A new generation of world simulation
• Google: We’re announcing Genie 4, our new world model for real-time interactive simulation
• Google DeepMind on X: the Genie 4 announcement
• Sundar Pichai on X about Genie 4

The Google and Google DeepMind blog posts are dated 27 October 2026. Google DeepMind's post on X was published at 17:10 Oslo time the same day. All information in this article comes from these sources.

Illustration: AI-generated with Grok Bot / hogby.ai