Google’s Genie 4 builds worlds you steer with words, and its characters talk back

Genie 4 is Google DeepMind’s new world model, and the worlds it builds finally make noise. Describe a scene in a sentence, or hand it a reference image, and the model generates a place you can walk through with the keyboard while it plays the wind in the trees, the waves on the beach and the voices of the people standing there. Those people can now hold a conversation that reacts to what you do, which DeepMind describes as the true frontier of this release.

The bigger change for anyone who tried earlier versions is control. You can type an instruction that changes the weather, swings the camera or drops a new person or object into the scene, and Gemini rewrites your request so it fits what is already on screen. Between your commands, Gemini keeps steering the simulation so the world does not fall apart. It feels less like playing a video game and more like directing a film set that never stops rolling, which is how some early testers are already using it.

Gemini plans the scene while Genie paints it

Genie 4 leans on the rest of Google’s model family. According to DeepMind’s announcement, Gemini Omni anchors each simulation in real-world appearance and behavior, while other Gemini models handle the reasoning that keeps a scene coherent from one moment to the next. Gemini 3.8 Flash sits in the loop as a guiding agent: it watches the world as Genie 4 draws it, breaks a goal into steps, checks incoming frames to confirm each step worked and changes the plan when you interrupt.

DeepMind says the largest leap is in people. Characters move more naturally, respond to their surroundings and can be steered while the simulation runs, and the model handles physics such as aerodynamics and collisions. The demos on the company’s blog include a maple seed spinning down into the grass, a crowded rural train station in 1980s Japan and a first-person dragon race scored with thunderous wingbeats, all sound generated by Genie 4 itself. Genie 3 produced silent worlds at 720p and 24 frames per second; this time DeepMind has published no resolution or frame-rate figures, so the improvement can only be judged from its own clips.

Games are not the only target. DeepMind says simulating the world is a crucial step toward artificial general intelligence, and one demo pairs Genie 4 with Gemini 4 Argon to simulate Apptronik’s Apollo humanoid robot playing checkers. Argon splits the task into steps and feeds them to Genie 4 as text captions while the scene renders. Safe, endlessly variable practice worlds for robots and software agents are the long-term business case, long before anyone sells a Genie game.

Filmmakers and photojournalists got it first

DeepMind has been testing Genie 4 with groups drawn from the creative industries, journalism, education, gaming and engineering. Filmmakers use it as a draft mode for pre-visualization, stepping into a live scene and running several takes while they adjust camera angles, lighting, dialogue and action. Game developers go from an idea to an explorable prototype in seconds, before writing a line of engine code. Enterprise partners are trying it for quick scenario generation.

The most delicate experiment involves photojournalists, who are testing how a single archival photograph can become a simulated version of that place in seconds, populated and moving. DeepMind also mentions visualizing how a flash flood or a wildfire unfolds. A convincing walk through a real historical street, with invented people speaking in it, is also a convincing fake. DeepMind’s answer is SynthID, the invisible watermark it embeds in all video and audio Genie 4 produces, plus internal safety evaluations and a slow rollout. SynthID only helps if platforms and viewers actually run a detector, and nothing on screen tells a casual viewer the footage is synthetic.

The worlds still drift

DeepMind is unusually frank about the limits. Genie 4 holds an environment together over short sessions, but visual details and object permanence can shift as a simulation runs longer, after several complex prompt edits or as you move around the scene. Complex physical interactions, accurate real-time audio, legible text, spatial accuracy and long-term consistency are all listed as unsolved research problems. Put simply, the bookshelf you passed a minute ago may not be the same bookshelf when you turn back.

There is also no way for most people to try it yet. Genie 4 is a research preview limited to trusted testers. Google says it is coming soon to Google AI Ultra subscribers through Project Genie, its web app for generating worlds, and to early access partners of the Gemini API and Google AI Studio, with a waitlist already open. There is no release date, no price beyond the Ultra subscription, Google’s most expensive AI plan, and no word on which countries come first. When Project Genie launched on January 29, 2026, running Genie 3, it was open only to Ultra subscribers in the United States aged 18 and over, and sessions were capped at 60 seconds.

DeepMind announced Genie 4 on October 27, 2026, in a post signed by Raia Hadsell, Diego Rivas and Hang Qi. The line began in February 2024 with a model that turned images into simple 2D platform games, moved to 3D with Genie 2 in December 2024 and reached real-time exploration with Genie 3 in August 2025. Each generation added something: a third dimension, then real-time movement, now sound and speech. What Genie 4 still has to prove is that a world can remember itself for longer than a coffee break.

Tags: artificial intelligence, Google, Gemini, Google DeepMind, World models, Genie 4