Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering

Google has released Genie 4, a significant update to its generative world model series, introducing audio as a new output modality alongside more realistic environments and the ability for users to steer generated scenes in real time. The model can now sustain a consistent scene for up to two minutes - double the limit of its predecessor, Genie 3 - addressing one of the more practical constraints of earlier versions.

The addition of audio marks a meaningful expansion of what world models can produce. Previous iterations of Genie focused on visual generation, but ambient sound and synchronized audio feedback are important components of believable interactive environments. Incorporating audio brings Genie 4 closer to being a self-contained simulation layer, rather than a purely visual one.

Physics fidelity and character quality have also improved noticeably. More humanlike characters and better-behaved object interactions are central to making generated worlds feel grounded rather than abstract. These improvements matter particularly for potential applications in game prototyping, interactive media, and training environments for embodied AI agents - areas where Genie has been positioned as a research and development tool since the original model launched.

Real-time steering is perhaps the most practically interesting addition for developers and researchers. The ability to influence a scene as it unfolds - rather than generating a fixed output and starting over - opens up more dynamic use cases and brings the experience closer to genuine interactive simulation. Whether Genie 4 will be made broadly available or remain within Google's research infrastructure is a detail to watch, as access has historically been limited for this model family.