IS YOUR CODEX
HAVING A MOMENT?
A pelican. A bicycle. What could possibly go wrong? See what rides and what falls apart.
One correct. Two incorrect.
Editorial examples, not model scores. Prompts and run conditions may differ.
Incorrect result

Incorrect result

Sources & original HTML
Editorial examples, not model scores. Prompts and run conditions may differ.
The reference shows the requested model, not independently verified. Contributed metadata is supplied by the creator.
Fresh from the flock.
See what everyone else’s pelicans look like.
gpt-6.1-sol (medium)
Pelican test
gpt-6-astra (medium)
Pelican test
gpt-6-astra (medium)
Pelican test
GPT-6.1 Sol (low)
鹈鹕侠
gpt-6-astra (ultra)
Pelican test
gpt-6.1-sol-max
My pelican ride
SOL6.1 (xhigh)
奶龙服海鸥骑自行车
gpt-6-astra (ultra)
Pelican test
GPT-6-Astral Xhigh
My pelican ride
GPT-6-Sol Xhigh
My pelican ride
GPT6-Luna medium
我的鹈鹕骑行
gpt-6-sol (medium)
Pelican test
gpt-6-astra (medium)
Pelican test
gpt-5.6-sol (medium)
Pelican test
gpt-6-astra (medium)
Pelican test
gpt-6-astra (medium)
My Pelican
Codex (model not reported) (medium)
Pelican test
GREAT EXPECTATIONS.
INTERESTING RESULTS.
The good, the weird, and the “I pay for this?” Find your spirit pelican. Send it to a friend.
Public benchmark examples and reader-contributed work. These are curated examples, not live test reports.
Your agent can do better?
Copy a prompt, run it locally, and bring your first result back.
YOUR TURN,
PICASSO.
Three ways to try it. Choose just one.
Paste a prompt
Copy this prompt into your coding agent and keep the first result.
Do not load persistent memories or call subagents. Create an HTML file containing a 2D SVG animation of a pelican riding a bicycle.
Output the complete, self-contained HTML source directly in your final response, without Markdown code fences or explanations.
Do not call tools, read files, load skills, access the network, or write files. The caller will save your final response.
All visible text in the animation, including titles, labels, and captions, must be in English. Set the HTML document's lang attribute to en.Public benchmark · pelican-bicycle-cli-v2-en
Run a test in your terminal
Connect this page to your local Codex, run once, then bring the HTML and score straight into a submission. Requires Node.js and an installed, signed-in Codex CLI.
npx --yes @ryqdev/pelican-test@latest connect --site https://codex-pulse.com/ --lang en --port 0Run the command above and keep the terminal open. Open its private pairing link within 5 minutes, then click Connect this computer on this page to confirm.
- Start terminal
- Connect browser
- Run test
- Review result
Your local test
Connect to the terminal you just started. One run uses your local Codex account. Keep that terminal open.
Waiting for your terminal
Run the command above and keep the terminal open. Open its private pairing link within 5 minutes, then click Connect this computer on this page to confirm.
Install a skill for repeat tests
Install for your coding agent, then ask it to use codex-pulse. The installer lets you choose your agent; generating with run currently requires local Codex.
npx skills add ryqdev/codex-pulse-skill --skill codex-pulse -gMade one? Upload your drawing with its model and prompt. Published posts get a link and a share card.
Post your pelicanLESS DOOMSCROLL.
MORE PELICAN.
The story behind the promptWhen your AI seems off, it helps to have something you can look at. Simon Willison’s pelican test starts with one request: draw a pelican riding a bicycle.
Does a bad pelican mean Codex is degraded?
No. One drawing cannot establish a quality regression or account-level restriction. Prompting, model settings and randomness all matter. This launch is a place to compare and share examples, with no live quality verdict.
Can I upload my own result?
Yes. Sign in, upload an SVG, PNG, JPEG or WebP, and add the model and prompt. New submissions go into review. You can view or delete your own work from My posts; published drawings get a public link and share card.
What do the gallery labels mean?
They are playful editorial labels, not benchmark scores. The starter artworks come from the linked public collection and keep their original model names. They are not represented as current Codex generations.






