Google has turned “generating playable worlds on the fly” into an API. On November 19, the world model Genie 4 officially opened to developers, callable through the Gemini API’s multi-turn interaction interface and AI Studio; the day before, the consumer-facing Project Genie had already been unlocked for AI Ultra users.

Press a Key, and the World Is Generated Instantly
A single line of text, a reference image, or a mix of text and images can generate a continuous interactive world at 720p and 24fps, with ambient sound and dialogue generated in sync, and all outputs are embedded with an invisible SynthID watermark. The on-screen response to real-time keyboard control is about 0.2 to 0.5 seconds, and typing text can also dynamically change the weather, lighting, objects, and even story events. Gemini inference keeps the world state coherent. It is positioned for interactive prototyping, spatial visualization, and robotics and agent training. Its shortcomings are also blunt: only 16:9 landscape; frame-by-frame rolling generation; no persistent 3D assets; once a session ends, it cannot be resumed; it stutters when network round-trip latency exceeds 150 milliseconds; and it is currently served only from U.S. data centers.
No Free Tier, and Closing the Page Doesn’t Stop the Meter
Billing is token-based: text and image input at $1.50 per million tokens, text output at $9, and video output with stereo at $17.50, which works out to about $0.10 per second, or $6.05 per minute.
The official guidance also flags two pitfalls: the visuals are most coherent during the first 45 to 90 seconds, and a single session is capped at 120 seconds; simply closing the webpage does not stop the meter—you must explicitly call complete() or cancel() to stop the charges. During the same period, Google also launched the reasoning model Gemini 4 Neon and a Robotics plugin.