Google DeepMind has unveiled Genie 4, a model that generates a photorealistic world while you move through it, with sound and characters who talk back. You steer with keyboard controls and typed prompts, and the scene changes as it runs. For now, only trusted testers have access.
According to the Google DeepMind blog, Genie 4 generates its own audio, which DeepMind calls a first for its real-time world models. Each scene carries the sounds of its setting, from wind in trees to waves on a beach, and characters hold conversations that shift with what the user does. Typing a new instruction mid-scene (a change in weather, a new person, a moved camera) works too, and Gemini rewrites the prompt so it fits the surrounding scene.
Gemini does more than translate prompts. DeepMind says Gemini 3.8 Flash runs inside the loop as a supervisor: it watches incoming frames, splits a goal into steps, checks that each step actually happened, and revises the plan when the user interferes. That supervision is how the lab claims to get reliable cause and effect over several steps.
The most interesting demo points away from entertainment. DeepMind paired Genie 4 with Gemini 4 Argon, which breaks a task into steps and feeds them to Genie as text captions while the simulation renders. The example shows Apptronik’s Apollo humanoid robot playing checkers. A simulated robot in a simulated room is a cheap place to find out how an AI system fails before it touches hardware, though the post offers no measure of how well results carry over to a real machine.
Testers are already trying other uses. Filmmakers step into a live scene and run several takes, changing camera angle, lighting, and dialogue on the fly. Photojournalists turn one archival photograph into a simulation of that place. Game developers go from an idea to an explorable prototype before writing any engine code, and enterprise partners are evaluating it for quickly generating test scenarios.
DeepMind is open about the weak spots. Visual details and the permanence of objects can drift, most of all in long sessions, after piles of intricate prompt edits, or when the user roams around. Readable text, accurate spatial layout, and staying consistent over time are still open problems. The post includes no benchmark scores and no independent evaluation, so the quality claims rest on the lab’s own recordings.
On safety, Google says Genie 4 went through internal evaluations and follows the same policies as its other models. Every video and audio output carries SynthID, the company’s invisible watermark for AI-generated content. That matters for a model whose output looks like live footage with synchronized speech.
Access comes in stages. Subscribers to the Google AI Ultra plan come first, inside Project Genie, followed by early partners using Google’s developer tools, and a waitlist is open. The post gives no date beyond “coming soon.”
Google has not said how many minutes a world holds together or what API access will cost. Those two answers decide whether game studios and robotics teams can build on Genie 4 or only watch the demos.
Reported by the Google DeepMind blog on 27 October 2026.
