Google's Genie 4 makes its own soundtrack. Access is the catch

Google DeepMind announced Genie 4 on October 27, a world model that generates sound as it generates the scene. There is no benchmark, no parameter count, no price and no public access. The only evidence is a set of recordings and Google's own published prompts.

The architectural claim worth taking seriously

DeepMind calls Genie 4 its "first real-time world model to generate audio natively", with music, ambient noise and interactive dialogue produced alongside the frames rather than dubbed on afterwards. That distinction is real. A post-hoc audio track can be cut to a finished clip; joint generation has to commit to a soundscape while the physics is still undecided. Gemini Omni anchors each simulation in real-world appearance and behaviour, while other Gemini models supply the reasoning that keeps a scene continuous. Steering runs through language: move the camera, change the weather, add a person, and Gemini rewrites the request to fit what is already on screen, then keeps the world progressing between inputs.

The demos are authored, and the prompts show it

The page publishes the prompt text behind its scenes, which is the most revealing part. The lakeside saxophone shot begins "Use the reference image as the exact first frame and visual basis for the scene," then specifies a man in a loose brown suit, "photographed on 16mm film with subtle grain," and "absolutely no HDR, no dramatic contrast." A 1980s Japanese station scene closes with prohibitions: "No empty platforms, no distant train, no sparse crowds, no centered symmetrical composition." This is a shot list, not improvisation. Each of the four capabilities is a curated recording with a prompt written to make it land, in a two-minute film that states every accompanying sound came directly from the model.

What a world model is for, according to Google

Google frames world simulation as "a crucial aspect of building AGI" and the outputs as safe training environments for robots and agents. TechCrunch reports a Gemini Robotics experiment in which Gemini 4 Argon broke tasks into steps that Genie 4 simulated, and notes the consistency window has doubled to two minutes. Its engineers also concede the gaps on the record: details drift as a simulation runs, audio accuracy is unsolved, and in-world text is not reliably legible.

Access is a research preview at best. Trusted testers now, a waitlist, then AI Ultra subscribers in Project Genie and early partners on the Gemini API and AI Studio.

What would settle it

Independent hands on the model, a fixed protocol, and every scene's prompt published rather than one. Whether the audio truly follows the picture, rather than being reseeded per shot, is checkable, and nobody outside Google has checked it. Until then, Genie 4 is a demonstration of a direction with a genuinely new soundtrack.