Genie 4 adds sound and dialogue to interactive worlds in limited preview

Google DeepMind began a Genie 4 research preview for trusted testers on October 27. Its new world model generates music, ambient sound and dialogue as users explore and reshape simulated scenes.

Testers navigate with a keyboard and mouse, using written prompts to change weather, reposition the camera or add people and objects. Gemini adapts each instruction to the scene and steers events between user inputs.

Real-time exploration and text-driven world events appeared in the earlier research release. DeepMind introduced Genie 3 on August 5, 2025, describing 720p worlds at 24 frames per second, consistency for several minutes and visual memory extending back up to a minute. Weather changes and new objects or characters were already demonstrated. Those figures describe Genie 3.

The Genie 4 launch film describes turning a picture into an explorable environment and presents characters responding in dialogue. One warns another about a trap after a suggestion to attack a castle. The narration presents virtual robot practice as a step toward possible future physical-world work.

In an early robotics example, Gemini 4 Argon passes a multistep plan as text to Genie 4 while it renders the simulation.

In a developer discussion, the team distinguishes playback frame rate from the delay after a user action. Smoother imagery and feeding input back into generation impose separate demands. Combining generated imagery with code that preserves world state remains a research direction.

The discussion also shows a development interface that can pause a simulation. The team explains that typing while the stream keeps moving makes it hard to place instructions at the right moment. Whether pausing becomes a user-facing capability remains open to testing and feedback; queuing instructions ahead of time and inserting them during a session are also being considered.

Preview sessions and existing public access

Genie 4 supports a few minutes of interaction. Details and object permanence can drift during long sessions, complex edits or navigation. Accurate geography, multi-agent interactions, legible text and real-time audio remain research challenges.

Project Genie, the subscriber-facing web app, is a separate product stage. Google introduced it on January 29, 2026 as a Genie 3 prototype for sketching, exploring and remixing worlds. That announcement explicitly said the app did not yet include some research capabilities, including promptable events during exploration. That January rollout initially covered U.S. Ultra subscribers aged 18 and over.

Current Project Genie help requires a personal Google Account, age 18 or older and an Ultra subscription where the plan is available. Those without access can still view the gallery of user-created worlds. The help page says world generation in this prototype does not consume AI credits. The page, which describes Genie 3 limitations, allows 60 seconds of exploration and then regeneration or video download. Reusing a prompt can produce different results. That timer documents the existing app, rather than establishing Genie 4 rollout terms.

Genie 4 video and audio will carry SynthID. Its documentation describes an imperceptible signal embedded in content for a detector to identify. Google designs visual markings to withstand changes such as cropping, filters and lossy compression, while audio markings are inaudible.

Broader Genie 4 access is planned for Ultra subscribers in Project Genie and early Gemini API and AI Studio partners. No specific date or price has been announced.