Skip to content
VibeCodeBench

The Flan Test

Claude Opus 5.5 made this.

AnthropicEffort: High, level 3 of 5

Try this: Tap Shake the plate, then scoop it with the Spoon.

Made by a Claude Code agent. It wrote the file in one go and never ran or fixed it.

The Flan Test by Claude Opus 5.5: a still of the demo
works · try 1 of 2
Click it to play here.Playing here. Done gives the page back.

Receipt

Made in
Claude Code
Paid with
Claude subscription
Time
32 min 2 s
Tries
Try 1 of 2
Date
Sep 30, 2026
By
a Claude Code agent

Made on a regular paid subscription, not the API.

Tries

  1. Try 1works32 min 2 s · 33.9 KBshownCode of try 1
  2. Try 2works17 min 54 s · 53.1 KBCode of try 2
How it was made

Each try was made the same way: a Claude Code subagent running Claude Opus 5.5 at high effort got only the prompt, pasted as is, and wrote index.html in a single step, without opening or testing it. Nobody edited the files.

Why there are two tries: on Oct 1, 2026 the build workflow restarted after a session limit and ran this prompt a second time, overwriting the first file. Before the site was published, its data listed only try 2, as if it were the only one. Try 1 was recovered byte for byte from the session logs on Nov 11, 2026.

  • Try 1: Recovered byte for byte from the session logs on Nov 11, 2026. Checked that day in a headless browser: no console errors, renders on desktop and phone. Its time runs until the file was written: the account session limit cut the agent before its last message.
  • Try 2: Accidental second try (the workflow restart).

How it was checked

Console errors
none
Links and libraries
all allowed
Hands-free demo
yes
Phone check
390x844
Libraries it loads
cdn.jsdelivr.net/npm/[email protected]/build/three.min.js

What the AI said

Nothing about what it built.

The agent wrote the file and was cut off by the account session limit before it could describe it.

See the code

The benchmark

The Flan Test

Poke, stretch and scoop a 3D caramel flan with real soft-body physics.

I keep the exact prompts private for now.

Every AI on this benchmark