Google DeepMind’s Genie 4 builds AI worlds you can talk to, and the people in them talk back

Google DeepMind’s playable AI worlds now have sound, and the people in them can talk back. The lab unveiled Genie 4 on Tuesday, a world model that generates music, background noise and spoken dialogue in real time alongside the world you’re moving through, and lets you change the scene by typing while it runs.

For now, almost nobody can try it. Genie 4 is a research preview for a small group of trusted testers, with a waitlist open in Project Genie, Google’s app for making and exploring these worlds.

What’s new in Genie 4

Genie turns a prompt or an image into a world you can walk around with your keyboard. The last version, Genie 3, was previewed in August 2025 and has powered Project Genie for Google AI Ultra subscribers in the US since January, in silent clips capped at 60 seconds.

DeepMind calls Genie 4 “natively multimodal”, and its post sets out four main advances:

• Sound: it is the lab’s first real-time world model to generate its own audio, from wind in the trees to waves on a beach. DeepMind calls human speech “the true frontier”: characters hold conversations and react to what you do, “all in real-time”.
• People and physics: DeepMind says simulating people is “the most dramatic leap”, with characters that “move naturally” and respond to their surroundings. It also handles physics such as aerodynamics and collisions.
• Live edits: you can type changes as the world runs, such as new weather, a different camera angle or an extra person or object, and Gemini adapts your prompt to fit the scene. Changing a world mid-exploration was something Google said Project Genie couldn’t do yet when it launched.
• Events that play out: set something in motion and Genie 4 simulates what happens next, including how the people in the scene move, react and speak.

Gemini runs the show behind the scenes

Genie 4 leans heavily on Google’s Gemini models. Gemini Omni “anchors the simulations in real-world appearance and behavior”, DeepMind says, while Gemini 3.8 Flash acts as “a guiding agent in the loop”: it watches the frames as they arrive, breaks a goal into steps, checks each step worked and changes its plan if the user steps in.

The lab has also paired Genie 4 with Gemini 4 Argon, its most capable model, for its robotics research. In one demo, Argon splits a game of checkers into steps and feeds them to Genie 4 as captions, so the world shows Apptronik’s Apollo humanoid robot playing it out. DeepMind describes simulating the world as “a crucial aspect of building AGI”, and pitches Genie 4 for building training worlds for robots and AI agents.

Who is testing it, and what still goes wrong

DeepMind says testers in film, journalism, education, gaming and engineering have been trying it. Filmmakers run several “takes” of a scene; photojournalists are turning a single archival photo into a simulated scene “in seconds”; and game developers are building playable prototypes “before writing a single line of engine code”. Every video and audio clip Genie 4 makes will carry Google’s invisible SynthID watermark, which anyone can now check with Google’s free detector.

The lab is open about the limits. Genie 4 keeps a world coherent “over short sessions”, but details and object permanence “can still shift or drift”, especially in longer runs, after several complex edits or as you move around. Complex physics, audio accuracy, legible text, spatial accuracy and long-term consistency are all still being worked on. The post gives no figures for session length, resolution or frame rate.

Some familiar names are in the credits. Jack Parker-Holder, Matt McGill and Philip Ball are listed among Genie 4’s core contributors: they are the former DeepMind researchers behind Emulate, the London start-up that was in talks to raise up to $700 million in September.

When you can try Genie 4

After the testers, Genie 4 goes to Google AI Ultra subscribers in Project Genie, then to early access partners on the Gemini API and Google AI Studio. DeepMind says it is “coming soon” but gives no date or price. Project Genie is currently open only to Ultra subscribers aged 18 and over in the US, and Google hasn’t said when Genie 4 will reach the UK, Canada, the EU, Australia or anywhere else. Until then, Google’s Playground turns a sentence into a game you can share.

Why it matters

Sound and talking characters push world models from silent tech demos towards something closer to interactive film and games, the prospect that sent video game shares lower when Project Genie launched. But Genie 4 is still a preview with no date, no price and no numbers on how long its worlds hold together, so how much of the demo reaches paying users is an open question.

Sources: Google DeepMind (Genie 4 announcement); Google (Project Genie launch)