Google DeepMind Genie 4: real-time world model with native audio

On 27 October 2026, Google DeepMind unveiled Genie 4, the latest iteration of its world models, during a limited research preview for trusted testers. AI Business recently delved into the impact of this system, which can generate audio simultaneously with simulated environments. This real-time world model can create ambient sounds, music, and character dialogues in the same pass as the visuals, rather than adding them afterward.

The system functions by building interactive worlds frame by frame in real time, allowing users to control scenes through natural-language prompts. Operators can adjust weather conditions, move the camera, introduce new characters or objects, and the model maintains the world's operation while applying each instruction. Gemini 3.8 Flash operates alongside Genie 4 as a guiding agent, breaking down the requested scenario into steps, monitoring the generated frames, and adapting the plan when the result deviates from the intended outcome.

In a demo, a simulated humanoid robot, Apptronik Apollo, played checkers inside the generated environment, illustrating how embodied agents can operate in worlds that did not exist prior to the prompt. Genie 4 builds upon Genie 3, incorporating more realistic physics, such as aerodynamics and collisions, as well as more natural human movement. Each output includes a SynthID watermark, indicating that the generated material originates from the system.

Google AI Ultra subscribers will access the system via Project Genie, while early-access partners will integrate the model through the Gemini API and Google AI Studio. However, no pricing information or rollout dates have been announced. The DeepMind report also highlights the system's limitations, such as visual drift in extended sessions and approximate physics. Text accuracy and long-term consistency are also identified as open problems.

The potential enterprise applications of Genie 4 include using generated worlds as rehearsal environments for robot and agent behavior testing before real deployment. This could replace some of the costly work typically conducted on physical prototypes. However, present accuracy levels limit this use case, as approximations deemed acceptable for entertainment may not be suitable for engineering validation.

The practical value of Genie 4 will depend on advancements in fidelity and session stability from the research preview stage to the planned rollout.