Real-time world model Genie 4 generates its own audio

Google DeepMind on October 27 announced Genie 4, a world model that generates an interactive scene in real time along with its ambient sound and its characters’ dialogue, and takes text prompts that change the scene while it runs. Outside the trusted testers who have it as a research preview, the way in for now is a waitlist: Genie 4 is coming “soon” to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, with no date given.

A world model predicts how an environment evolves and how actions change it, generating the path ahead as the user moves instead of serving a fixed 3D scene. In the recordings Google DeepMind published, users move and turn the camera with standard keyboard controls and type prompts to steer what happens next.

Getting in

The Genie 4 waitlist runs through Project Genie (sign-up form). Project Genie is the prototype web app Google opened to Ultra subscribers on January 29, at the time powered by Genie 3, Nano Banana Pro and Gemini. That rollout covered U.S. subscribers aged 18 and over and was billed as an experimental research prototype.

Genie 3 travelled the same route. Introduced in August 2025, it first went to trusted testers, and Google described the Project Genie launch as the next step to broaden access. For Genie 3, the gap between the trusted-tester preview and paying subscribers was five to six months. On the developer side, Genie 4’s first stop is early access partners of the Gemini API and Google AI Studio.

What Gemini does in the loop

Genie 4 builds on the Gemini family. Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a simulation coherent and continuous. The real-time piece is Gemini 3.8 Flash, running as a guiding agent: as Genie 4 generates the world, it watches the scene, breaks a goal into a sequence of steps, reviews incoming frames to check that each step succeeded, and changes its plan when the world shifts or the user intervenes. Between user inputs it also steers the scene so the world keeps moving forward, and when a user asks for a change (different weather, a camera move, a new person or object), Gemini adapts the prompt to fit the surrounding scene.

Audio is new to the line. Google DeepMind calls Genie 4 its first real-time world model to generate audio natively: ambient sound such as wind in trees or waves at a beach, music, and dialogue in which characters react to the user’s inputs or to events in the world. Every sound in the published demos came from Genie 4 itself, the company says. On the visual side, it claims more realistic physics, citing aerodynamics and collisions, and people who move naturally and can be steered in real time.

The same loop is being tried for robotics. Google DeepMind has been pairing Genie 4 with Gemini 4 Argon as part of its Gemini Robotics work: Argon splits a user’s task into steps and feeds them to Genie 4 as text captions while the simulation renders. The demo shows Apptronik’s Apollo humanoid playing checkers.

Who is testing it

Trusted-tester cohorts span creative industries, journalism, education, gaming and engineering. Game developers are using Genie 4 to go from an idea to an explorable prototype in seconds, testing visual styles, materials and interaction mechanics through prompts “before writing a single line of engine code”. Filmmakers run several “takes” of a scene in a live simulation, adjusting camera angles, lighting, dialogue and action. Photojournalists are turning a single archival photograph into a simulated environment, and enterprise partners are evaluating it for what Google DeepMind calls rapid “zero-to-one” scenario generation.

Where it drifts

Google DeepMind still calls Genie 4 an early, experimental model. It keeps environments coherent over short sessions, but visual details and object permanence, meaning whether things stay put when out of view, can shift as a simulation runs longer, after multiple complex prompt edits, or while the user navigates the scene. The prompt editing that headlines the launch is itself on that list of triggers. For comparison, Genie 3 worked with a one-minute memory at 720p and 24 frames per second.

Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency all remain open research areas, by the company’s account. All video and audio Genie 4 generates will carry SynthID, Google’s imperceptible watermark for AI-generated content, and the model went through internal safety evaluations before the preview.

Google DeepMind presents the Gemini pairing as an early step toward hybrid world models, in which “Gemini works behind the scenes to map the structure and plan the narrative timeline, while Genie generates the immediate physical reality”.