Google DeepMind announced Genie 4 on October 27, and the short version is that its world model now makes noise. Genie 4 generates an interactive environment in real time that you move through with a keyboard and steer with text prompts, and for the first time it produces the audio alongside the picture: ambient sound, music and spoken dialogue. DeepMind calls it "our first real-time world model to generate audio natively". It is a research preview for trusted testers today, with access promised to Google AI Ultra subscribers and API partners in the coming weeks.
What changed from Genie 3 is mostly about how long a world holds together and how much you can do to it once it exists. TechCrunch's report puts consistent scenes at up to two minutes, double Genie 3's limit, and DeepMind's post lists more photorealistic visuals, simulated physics such as aerodynamics and collisions, and characters that move and behave more naturally. The headline interaction is mid-scene editing: you can change the weather, move the camera, or drop in people and objects while the simulation runs, and Gemini adapts the request to fit the scene. TechCrunch describes Gemini acting as a planner that breaks tasks into steps and checks frames while Genie renders them, which Google's research director Tejas Kulkarni framed as the shape of things to come: "What we'll increasingly see in the future are these hybrid world models."
Where game makers come in
The games angle is not incidental. DeepMind says game developers and creative technologists are among the testers, using Genie 4 to turn an idea into an explorable prototype and to try visual styles and mechanics through prompts before writing any engine code. That is a narrower claim than "AI makes games", and it is worth keeping narrow. What Genie produces is a short, steerable simulation, not a game with rules, saves or a build you can ship. The plausible use is the one DeepMind names: a designer describing a level, walking through a rough version of it within minutes, and deciding whether the idea is worth building properly. TechCrunch adds film prototyping and storyboarding, and robotics simulation, as other early uses.
The limitations are spelled out by DeepMind itself rather than discovered by critics, which is to its credit. Visuals and object permanence drift over longer runs. Complex physical interactions are unreliable. Audio accuracy is not solved, and in-world text is often illegible. Spatial accuracy and long-term consistency remain open problems. DeepMind engineer Diego Rivas told TechCrunch it is "still early days for Genie 4" and that the end-to-end product experience is unclear. Everything Genie 4 generates, video and audio alike, carries Google's SynthID watermark.
Why it matters
World models are the part of the AI race closest to games, and the field is consolidating fast: TechCrunch notes AMD's acquisition of World Labs, Amazon's reported talks to buy Decart and Runway's GWM Worlds 2 from a couple of months ago. Genie 4 adding sound and live editing moves the technology from a silent tech demo towards something a person could actually sketch with. For players, nothing changes yet; there is no Genie-made game to buy and no sign of one. For developers, especially small teams who prototype in grey boxes, the question becomes whether a two-minute steerable sketch is useful enough to change how ideas are tested, and the answer depends on access more than on the demo.
That access is the next step, and Google has given it only in outline. Trusted testers have it now. Google AI Ultra subscribers will get it through Project Genie, Google's tool for generating worlds from text or image prompts, and early-access partners through the Gemini API and Google AI Studio, both described as coming in the next few weeks. A waitlist for Project Genie is open. Google has not announced pricing beyond the Ultra subscription or a date for wider availability.
Reporting notes