Crane Arena

Lab

My son is five and a little obsessed with tower cranes. So I gave four AI agents the same 20 KB prompt: build a tower crane construction site in Three.js, in the browser, in one shot, with no help and no feedback. Each agent worked alone in an empty folder, in its own native tool. These are the four results, running live, plus the prompt they all got.

UI labels translated to English, code otherwise as the agents wrote it. Scenes load Three.js from a CDN and run straight in the browser. Desktop with a mouse works best.

Controls

  • Mouse drags the camera, wheel zooms.
  • Space runs the construction.
  • N and D switch night and day.
  • Keys 1 to 5 jump between cameras.
  • H hides the panel.

Each agent wired its own panel, so the controls differ a little between scenes.

The four cranes

Crane A by Qwen3.8-27B: construction site after three repair rounds
AQwen3.8-27B on Cerebras

The first run stopped mid-file at line 987, cut by the output limit, and the page was blank. After three repair rounds the scene runs: site, crane, trucks, workers. The house still does not grow, and it draws 78,125 calls per frame where the prompt asked for under 300. Then the $10 of prepaid Cerebras credit was gone.

Open live scene
Crane B by Claude Opus 5.5: construction site with working interface
BClaude Opus 5.5

The strongest of the four. Every detail worked out, good light and materials, and the interface actually works. 57 minutes, 308K output tokens, about $13.84 at API list prices. I judged the scenes blind and guessed this one was Gemini Argon.

Open live scene
Crane C by Gemini 4 Argon: clean construction site scene
CGemini 4 Argon

Second place, and the fastest of the three that finished: 28 minutes, 102K output tokens, about $1.71 at its launch price. A bit less detail than B, but clean and solid. Blind, I was sure this was Opus.

Open live scene
Crane D by Claude Haiku 5.5: construction site with floating workers
DClaude Haiku 5.5

A complete site that runs, with one tell: some workers hang in the air. 52 minutes, 439K output tokens, about $2.13. For a small model on a 20 KB spec this is further than I expected it to get.

Open live scene

And one for fun

Not part of the race. After the race I asked for a side scene: the crane drops a flying roadster and it lands on its thrusters. R replays it, O frees the camera.

Open the roadster scene
The prompt

One file, 20 KB, the same for all four: tech stack, the crane, the building, workers and vehicles, day and night, a performance budget of under 300 draw calls. No questions allowed, one pass, stop when done.

Read the prompt
The video

The race, all four cranes side by side, my blind guesses and the full receipt of time and cost.

Watch on YouTube
The write-up

How I judged them blind, what I got wrong, and how I pick a model now.

Read the post

Українська версія