DeepSeek V4.1 Flash built this three.js day-to-night scene from one prompt: 2h 22m, 305 calls, about $1.16

I gave DeepSeek V4.1 Flash one prompt and let it run. It came back 2h 22m later with this. The brief: "Golden Hour", a three.js scene of a high-rise bedroom where one time-of-day control moves the sun from morning to night. No downloaded models, textures or HDRIs, everything built in code. Full prompt in the comments. What surprised me was how it worked. It didn't just write the code and stop. It served the page, took screenshots at 07:00, 12:00, 18:30 and 21:30, looked at them, fixed what looked off, and went again. 305 model calls, no approvals from me. Then it tested it like a QA person would: Chrome desktop and a phone-size viewport with no console errors, every control working (clock, slider, arrow keys, day loop), a Firefox smoke test, and frame rates (86 fps at 1080p, about 56 at 2x pixel ratio). And it finished with an honest list of what it couldn't nail: - it's stylized low-poly, not photoreal - the window frames never cast the crisp shadow stripes I asked for, so golden hour is more of a soft haze - the black wardrobe reflects bronze at sunset - it couldn't test Safari Cost: about $1.16 at off-peak prices, with 99% of input tokens served from cache, at max reasoning effort the whole way. Disclosure: I work on AgenC Core , the open-source (MIT) terminal agent it ran in: github.com/tetsuo-ai/agenc-core