THE ESSENTIALS IN 30 SECONDS

  • Genie 4 generates video, ambient sound, music and dialogue together; DeepMind says it is the first real-time world model with native audio.
  • Users can change a running scene with text prompts, such as altering the weather or adding a character.
  • It is a research preview for trusted testers; wider access via Project Genie for Google AI Ultra subscribers and the Gemini API is planned.

Google DeepMind has unveiled Genie 4, the latest version of its "world model" — a system that generates an explorable, interactive environment in real time rather than a fixed video clip. The lab describes it as an early, experimental model, and access is restricted to trusted testers for now.

What is new

The main addition is sound. According to DeepMind, Genie 4 produces ambient audio, music and dialogue in the same pass as the visuals, and the lab calls it the first real-time world model to generate audio natively. Its predecessor, Genie 3, produced silent scenes.

Users move through a scene with standard keyboard controls and can type instructions while it runs — change the weather, move the camera, add a person or an object — without restarting. DeepMind says Gemini 3.8 Flash acts as a guiding agent that observes the scene, breaks goals into steps and adjusts as the world or the user's instructions change. The company also claims more convincing physics, including collisions and aerodynamics, and characters that react to their surroundings.

Who can use it

Genie 4 is a research preview. DeepMind says it plans to open it to Google AI Ultra subscribers through Project Genie and to early access partners on the Gemini API and Google AI Studio, with a waitlist available. No dates, regions or developer pricing have been announced. Startup Fortune notes that Project Genie previously required Google's top AI Ultra subscription.

DeepMind says trusted testers are using the model for film previsualization, turning archival photos into explorable scenes, prototyping games and building simulated scenarios for businesses. Generated video and audio carry SynthID watermarks.

Limits

DeepMind is open about weaknesses: visual details and object permanence can drift in long sessions, after several prompt edits or while navigating. Complex physical interactions, audio accuracy, legible text, spatial accuracy and long-term consistency remain research problems. The blog post also does not state session length, resolution or frame rate, making direct comparison with Genie 3 — which ran at 720p and 24 frames per second — difficult until independent testers report back.