token-maxxing.club

3 agents · 3 fresh Linux desktops · 1 prompt

I gave 3 AIs a blank Linux desktop and said: build a game.

Each row: the agent coding on its desktop, beside the game it shipped — then the games full size. Real times 53s · 686s · 870s, all on my Synthetic sub.

How each one did

Kimi-K2.6

53s

Synthetic · pi

53 seconds — fastest by a mile. 811 lines, a real playable game with rocks that split when you shoot them.

Its first build crashed on launch, so I gave it a clean re-run — same prompt, no fixes from me. Fast isn't always safe.

MiniMax-M3

686s

Synthetic · pi

Eleven minutes, 1,290 lines. The busiest HUD of the three — and it ran first try.

The steady middle: more code, more time, a solid game.

GLM-5.2

870s

Synthetic · pi

The maximalist this round — 1,360 lines, and the slowest at 14½ minutes.

Most code of the three. Waves, hyperspace, the works — it just took its time.

The three playable Asteroids games, side by side: Kimi-K2.6, MiniMax-M3 and GLM-5.2
What each shipped — captured mid-game. All three are real, playable Asteroids; all three split rocks when you shoot them.
Same prompt, three very different runs — and the same game out the other side. Kimi shipped in 53 seconds; GLM took fourteen minutes and the most code. More tokens and more time didn't mean more product. (Kimi's first build even crashed on launch — fast isn't free either.)

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

Join the waitlist →