Google Introduces New AI Model Genie 4 for Creating Interactive Worlds

Google DeepMind has introduced the AI model Genie 4, capable of creating interactive worlds with sound in real time. Developers have announced more realistic graphics, improved physics modeling, and character behavior.

Genie 4 belongs to the class of world models: such systems do not just create a sequence of frames but simulate environmental changes in response to user actions. A person can navigate the space using a keyboard and control the camera, while text commands can add objects and change the weather. Gemini adjusts requests to what is happening and monitors the world's sequence. According to TechCrunch, Genie 4 maintains a consistent scene for up to two minutes—twice as long as its predecessor.

One of the main changes is the built-in sound generation: the model creates noises, music, and speech simultaneously with the image. Characters engage in dialogues, react to user actions, and change their lines during the scene. Developers also announced improvements in aerodynamics, collisions, and other physical interactions, with the most notable step being the simulation of humans: characters move more naturally and respond to the environment. The demo features a hot air balloon flight over Antarctica, a Japanese station from the 1980s, dragon races, and a world of stained glass.

Genie 4 relies on Gemini models: Gemini Omni is responsible for the realistic appearance and behavior of objects, while Gemini 3.8 Flash acts as a managing agent—observing the scene, breaking down tasks into steps, and adjusting the plan when the user intervenes.

As part of Gemini Robotics, the model is being tested alongside Gemini 4 Argon: Argon breaks down tasks into steps, while Genie 4 models how the Apollo robot from Apptronik plays checkers. Such environments will help train robots before transitioning to the real world.

The model is already being tested in cinema, journalism, education, game development, and engineering. However, Genie 4 remains experimental: details and objects may change appearance over time, especially in long sessions. All video and audio will receive SynthID watermarks. Access will be opened to a limited group of testers, then to Google AI Ultra subscribers in Project Genie and partners through the Gemini API and Google AI Studio; the API cost has not been disclosed.