Genie 4 announced: Google is first opening its model to trusted testers

Google DeepMind announced its world model, Genie 4, on 27 October 2026. Designed to create interactive environments, the model lets users make changes with natural-language prompts while a scene is running and generate simulations with audio. Google is initially making Genie 4 available to trusted testers as part of a research preview. The company’s announcement also outlines a phased access plan through Project Genie for Google AI Ultra subscribers and early-access partners for Gemini API and Google AI Studio.

Scene flow changes with natural-language prompts

With Genie 4, users can use natural language to change the weather, turn the camera in a different direction, or add a person or object to the environment while a simulation is running. According to Google, the model interprets requests in the context of the surrounding scene. Even when users aren’t issuing new prompts, it tracks how the scene unfolds and helps events in the world progress consistently.

Google’s demo examples show different environments, characters, and camera movements. The company says the model can simulate physical events such as aerodynamics and collisions, and that characters can react to their surroundings and be directed in real time. These are statements and examples from the company’s announcement, not independent performance evaluations.

Native audio generation and Gemini support

One of the model’s standout features is its ability to generate sounds within the simulation to accompany the visual scene. Google says it can produce ambient sounds such as wind and water, as well as interactive conversations that adapt to events in the scene or user input.

In the system Google describes, Gemini models help the simulation maintain realism and continuity. Gemini Omni contributes to the environment’s appearance and behavior, while other Gemini models handle complex reasoning tasks. For real-time operation, Gemini 3.8 Flash monitors the scene, breaks goals into steps, checks progress, and updates its plan when the environment changes.

The Verge reports that experiences can last around two minutes, longer than the one-minute duration of the previous Genie 3 model. The publication also notes that details and object continuity may change as a simulation gets longer. Since these duration and limitation details aren’t included in Google’s announcement, they should be treated as information reported by the publication.

Workflows for architecture, interior design, and visualization

The announcement doesn’t describe Genie 4 as architectural modeling or rendering software. However, its approach to building interactive environments, modifying scenes with prompts, and generating ambient audio could be of interest for spatially exploring early design ideas or investigating how an environment might be experienced. Google says the model could be useful for applications such as film previsualization, prototyping game ideas, and robot training scenarios, but makes no specific claims about architectural workflows.

For studios, the key question is how these possibilities might fit into existing production pipelines. The announcement doesn’t provide information about integration with architectural software, file exchange, hardware requirements, or pricing. So it isn’t possible to conclude that the model will work directly with project files or replace existing rendering workflows. Once access is available, scene continuity, camera control, and whether design elements remain unchanged should be tested separately.

Access is limited for now

Google says it has begun evaluating Genie 4 with trusted testers as a research preview and is gathering feedback. Broader consumer and developer access is planned for later phases, and a waitlist for Project Genie is now open. The Verge reports that access for Google AI Ultra subscribers will be available soon.

This phase doesn’t mean Genie 4 is generally available. For architecture and visualization teams, the most prudent approach is to follow the product as a research preview until access conditions, pricing, hardware, and compatibility details are announced. If a trial becomes available, one key area to assess will be how well objects and spatial details are preserved in longer scenes.