On August 2, 2026, Andrej Karpathy shared on X that he gave Claude Opus 5 the first paragraph of The Lord of the Rings, with a budget of approximately 1 million tokens, or about $10, to render it as a 3D world using Three.js. Opus 5 ran for about 2 hours, generating approximately 5,500 lines of code and producing a 93-second animated video. The visuals are crude, but it successfully transformed text into a dynamic 3D scene. The point isn't that "AI spent $10 to create a finished game"—it didn't. Only a video was released, with no open-source code repository or playable version. The audio was not generated by a model; Karpathy said the narration was manually created using ElevenLabs because he cared about which voice was used. The real signal is that the "generation cost" of customized visual software is falling faster than AI's ability to inspect and experience its own work. Opus 5 cannot efficiently natively view videos or try out games; it can only slowly take screenshots across different nodes, during which multiple errors occur. Karpathy’s “GTA on demand” is a direction, not the current reality. Today, this is closer to an automatically generated browser-based stereoscopic box. Karpathy evaluated the result as "janky but fun." The model placed countless polygonal assets into 3D