token-maxxing.club

3 agents · 3 fresh Linux desktops · 1 prompt

I gave 3 AIs a blank desktop and said: build a landing page.

It starts on a bare desktop — each agent boots its dev server, then builds. The three terminals run down the left, the three pages down the right, then each finished page scrolls full size. Real times 249s · 424s · 616s, all on my Synthetic sub (Kimi via the Kimi CLI).

How each one did

Kimi 2.7

249s

Kimi CLI · K2.7 Code

Fastest by a mile — 4 minutes. 1,342 lines, the most of the three, and a genuinely premium page: glowing orbital hero, clean three-tier pricing.

The real Kimi 2.7, run through its own CLI. Quick and tasteful.

Open the page it built ↗

MiniMax-M3

424s

Synthetic · pi

Seven minutes, 1,255 lines. The busiest hero — a live code-editor mockup — and a punchy 'speed of neon' pitch. Ran first try.

The fewest lines, and arguably the most product per line.

Open the page it built ↗

GLM-5.2

616s

Synthetic · pi

Ten minutes, 1,305 lines. The 'speed of light' maximalist, with a glowing-planet hero and the most copy. The slowest of the three.

Its first run flopped — printed the page into chat instead of writing it. A clean re-run nailed it.

Open the page it built ↗
The three cyberpunk landing pages, side by side: Kimi 2.7, MiniMax-M3 and GLM-5.2
What each shipped — the hero of each page. All three are real, self-contained Astro pages built inline from scratch.
Same prompt, three very different-looking pages — and all three landed around 1,300 lines across a 2.5× spread in time. The fastest run shipped the most code. More tokens and more time didn't mean more product.

I run little races like this most weeks. The club is where they get shared first.

Join the waitlist →