Skip to content
VibeCodeBench

Swim the Kelp Garden

GPT-6.2 Sol made this.

OpenAIEffort: Max, level 5 of 5

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

Swim the Kelp Garden by GPT-6.2 Sol: a still of the demo
broken · auto-checked (Checked by an automated pass. Not hand-played yet.)
Click it to play here.Playing here. Done gives the page back.

Receipt

Made in
Codex
Paid with
ChatGPT subscription
Time
17 min 4 s
Tries
1 try · broken
Date
Nov 22, 2026
By
a Codex agent
API equivalent (estimate)tokens × list price
about $0.51

Estimate at list API prices. I used my regular paid plan, not the API.

Prompt and answer tokens at list price, thinking included; the coding tool's own instructions are not counted.

Try

  1. Try 1broken · auto-checked (Checked by an automated pass. Not hand-played yet.)17 min 4 s · 77.5 KBshownCode of try 1
How it was made

One-answer agent run: a Codex agent running GPT-6.2 Sol at max effort got only the prompt and wrote the file in a single write. It did not open or test it, and nobody edited it afterwards.

What it built, in its own words: I wrote index.html for a procedural underwater garden with swaying kelp, rocky arches, a wooden shipwreck, schooling fish, manta rays, and rising bubbles. The file includes touch and mouse swimming controls, collision boundaries, and a looping 20-second guided tour.

Limitations it reported: A JavaScript syntax error in the sand shading calculation prevents the scene from starting.

Status not checked by a person yet.

  • Try 1: Automated check: A JavaScript syntax error stops the scene from loading, so only an empty teal background with title text shows. There is no kelp, ship, fish or rays, and nothing responds to input.

How it was checked

Console errors
1
Links and libraries
all allowed
Hands-free demo
no
Phone check
390x844
Libraries it loads
cdn.jsdelivr.net/npm/[email protected]/build/three.min.js
See the code

The benchmark

Swim the Kelp Garden

A whole kelp forest under the sea, made only from code. Swim through it with one finger.

I keep the exact prompts private for now.

Every AI on this benchmark