3 agents · 3 fresh Linux desktops · 1 prompt
It starts on a bare desktop — each agent boots its Astro dev server, then builds live: the terminal up top, the page rendering below, one column per model. Then each finished site scrolls full size — the home page, then /destinations. Real times 187s · 297s · 416s, on my own subs (Kimi via the Kimi CLI, GLM on z.ai).
pi · z.ai
Under 3.5 minutes — the fastest, and the leanest at 932 lines. It still shipped the whole brief: an animated starfield hero on the home page and a full grid of worlds on /destinations, wired together by a real nav. The whole site, fastest, with the least code.
Its Synthetic endpoint returned empty completions all day, so this is the real GLM-5.2 on z.ai's native API. It won anyway.
pi · Synthetic
Five minutes, 1,275 lines — the most code of the three, and the prettiest single page: a moody 'Beyond Earth. Beyond imagination.' hero. But it stopped there and never built /destinations, so the multi-page brief came in half-done.
The most lines, the best-looking hero — and the only one that skipped page two.
Kimi CLI · K2.7 Code
Seven minutes, 1,261 lines — the slowest, but it left nothing out: a 'Journey Beyond the Stars' home page plus a real /destinations grid, sticky nav linking the two. The most finished of the three.
The real Kimi 2.7 through its own CLI. Took its time; built the complete site.
The fastest run also wrote the least code — and still built both pages. The one that wrote the most shipped only one. More tokens didn't mean more product.
I run little races like this most weeks. The club is where they get shared first.
Join the waitlist →