Google DeepMind announced Genie 4 on October 27, describing it as a world model that generates interactive, real-time environments rather than video clips. According to the company, the model is natively multimodal: ambient sound, music and interactive dialogue are generated together with the world, and all audio in its demos was produced by the model itself. It builds on the Gemini family, with Gemini Omni anchoring appearance and behavior while other Gemini models supply the reasoning that keeps a simulation coherent.
Users steer with standard keyboard and camera controls plus text prompts. A prompt can change the weather, move the camera or introduce a person or object, and Gemini adapts the instruction to the surrounding scene and keeps the world progressing between inputs. DeepMind says Genie 4 simulates aerodynamics and collisions and gives characters more natural movement, and that in multi-step tasks Gemini 3.8 Flash acts as a guiding agent that observes frames, breaks goals into steps and verifies whether each step succeeded.
Genie 4 is a research preview for trusted testers, starting with Google AI Ultra subscribers in Project Genie and early Gemini API and AI Studio partners. Generated video and audio carry SynthID watermarks. DeepMind lists remaining limits: object permanence and visual detail can drift in long sessions or after many prompt edits, and complex physical interactions, real-time audio accuracy, legible text and spatial accuracy are still unreliable.
Our takeThe interesting claim is not photorealism but the loop: a reasoning model watching generated frames and checking whether a step succeeded turns a world model into a testbed for agents. That is also where the weakness shows. The same page admits object permanence drifts, and a verifier judging generated frames inherits their errors. Until there is external ground truth, multi-step success is measured inside the simulation, not against the world.