Google previews Genie 4 with live audio and text-steered worlds

An interactive scene, not just a finished clip

Google has announced Genie 4, a model that generates interactive worlds and their sounds as a user explores them. Its October 27 announcement describes changing a scene through text instructions, including moving a camera, altering weather or adding characters.

Unlike a finished video, the scene can respond while it is running. Google says generated sound includes environmental noise, music and dialogue. These are the company’s descriptions of its system, not results from an independent product test.

Who can use it

Google says rollout has begun with a limited group of trusted testers. Access for Google AI Ultra subscribers through Project Genie, and early-access partners through the Gemini API and AI Studio, is still described as coming soon.

That distinction matters for anyone planning a workflow around the announcement. A research preview and a waitlist do not establish general availability. The announcement does not give a firm date for that wider release.

Steering the world does not guarantee its accuracy

DeepMind describes Gemini models helping direct the unfolding scene. In one setup, Gemini 3.8 Flash observes incoming frames, breaks a goal into steps and checks progress. The intended combination is a planning system guiding a generative simulation.

DeepMind also lists important limits. Object permanence—the ability to keep the same object consistent—can deteriorate as sessions grow longer or prompts become more complex. Physical interactions, sound accuracy, readable text, spatial accuracy and long-term consistency remain research problems.

Those caveats qualify the realism claim. An image can look persuasive while its objects change or its physical behavior is wrong. For engineering or robot training, visual quality alone cannot establish that a simulated result will hold outside the generated scene.

What remains to be demonstrated

Google presents creative prototyping and training environments as possible uses. The practical questions are how reliably a world follows an instruction, how long it stays consistent and whether outside users can reproduce the advertised behavior.

Model Current has not accessed the preview. This report verifies the announcement and its stated access limits; it does not independently validate the model’s performance.