Google DeepMind adds sound and interactive dialogue to Genie 4’s simulated worlds

Google DeepMind has introduced Genie 4, bringing sound to its AI-generated worlds and promising characters whose dialogue changes with the simulation. The update also targets more convincing physics and human movement. Google plans wider developer and consumer access, moving beyond a small testing group, but has not set a release date.

From a generated scene to a responsive world

Genie 4 follows Genie 3, which remained confined to the limited Project Genie web experience. DeepMind presents these systems as more than a way to make something that looks like a video game: it sees world models as a route toward artificial general intelligence and a tool for creating environments in which robots could train.

A world model generates an environment frame by frame as someone explores it. Instead of requiring a pre-rendered setting and separately programmed interactions, it builds the experience from a prompt in real time. DeepMind also points to rapid prototyping in creative work as a potential use. Those ambitions depend on how faithfully the generated world behaves, not simply how convincing one frame looks.

Gemini helps steer each frame

To keep the experience on course, Genie 4 uses Gemini 3.8 Flash as a guiding agent. It breaks down goals and evaluates output frames, helping the simulation preserve its original context as the user moves through different environments. The guiding model is meant to keep successive images connected to the requested experience rather than letting the scene drift.

Users can also change the scene while it runs, a feature carried over from Genie 3. They can alter the weather, introduce an element or move the camera, with Gemini helping steer each update. The interaction is therefore not limited to choosing an opening prompt and watching a fixed sequence play out.

DeepMind says the new version handles aerodynamics and collisions more realistically. It also says simulated people move more naturally and respond better to changes around them. These are company-described improvements, rather than an independent assessment of how accurately Genie 4 reproduces physical behavior.

Audio adds another layer. Genie 4 is DeepMind’s first world model with native sound generation, including ambient background noise and audio elements specified in a prompt. Beyond those sounds, DeepMind promises interactive human speech: dialogue that adapts when the simulation changes, rather than remaining a fixed accompaniment to the images.

The next step is wider access

A small group of trusted testers has been working with Genie 4 as DeepMind prepares a wider release. The company still describes world models as an active research area. Access is supposed to broaden soon, but interested users currently have a waitlist rather than a specific date for when they can begin.

• Developers: Google plans access through the Gemini API and AI Studio. Developer pricing has not been announced.
• Consumers: Access is planned in phases through Project Genie for Google AI Ultra subscribers. Ultra plans start at $100 a month.

Google may eventually extend consumer access to AI Pro and AI Plus, but that is a possibility, not a committed rollout. Project Genie will remain the consumer interface, with limits on how much users can do. For developers evaluating the planned release, both the price and the date remain open.