Semantics
Is it unmistakably a pelican riding a bicycle — not a duck, a stork, or two unrelated drawings?
5 pointsThe animated LLM benchmark · 08 runs · 27 Oct 2026
Eight AIs. One pelican.
One bicycle.
The pelican-on-a-bicycle test is the internet's quickest visual check for LLM drawing skills. We extended it: one prompt, eight models, and a scene that has to move.
Zero-shot · consumer clients · one run per model
Winner23.5/ 25
Doubao 2.1 ProThe only pelican that truly pedals.
- 08runs
- 25points
- 01prompt
- 00fixes
The exact prompt
Create an HTML page containing a 2D SVG animation of a pelican riding a bicycle. You don't need any tests, and don't add any restrictions.
Sent once, unchanged, to each model. No follow-ups, no fixes.
The field
The starting grid
Eight machines, eight interpretations, one shared task. Open a card to watch the full run.
Winner
23.5/25
ByteDance
Doubao 2.1 Pro
The only pelican that truly pedals.
Motion 4.5
Anatomy 4.0
Watch the run
21.5/25
Moonshot AI
Kimi K3.1 Pro
The prettiest ride in the race.
Motion 4.0
Anatomy 3.5
Watch the run
20.5/25
OpenAI
GPT-6.1 Sol
Inverse kinematics under a coral frame.
Motion 4.5
Anatomy 3.0
Watch the run
18.5/25
OpenAI
GPT-6.0 Luna
Lovely lake. Two frozen legs.
Motion 2.5
Anatomy 3.0
Watch the run
18.0/25
DeepSeek
DeepSeek Flash
A perfect bicycle that never pedals.
Motion 1.5
Anatomy 3.5
Watch the run
17.5/25
OpenAI
GPT-5.6 Sol
Beautiful coast. Runaway wheels.
Motion 1.5
Anatomy 3.5
Watch the run
16.0/25
Not stated
HY3
Happy pelican. Orbiting legs.
Motion 1.5
Anatomy 2.5
Watch the run
9.5/25
Agnes AI
Agnes 3.0 Flash
The rider floats behind the bike.
Motion 2.0
Anatomy 1.0
Watch the run
Scoreboard
The leaderboard
Editorial scores from one zero-shot run. The total is out of 25.
| Rank | Bib | Model | Semantics | Anatomy | Bicycle | Motion | Polish | Total |
|---|---|---|---|---|---|---|---|---|
| 1 | 03 | Doubao 2.1 ProByteDance | 5.0 | 4.0 | 5.0 | 4.5 | 5.0 | 23.5/25 |
| 2 | 02 | Kimi K3.1 ProMoonshot AI | 5.0 | 3.5 | 4.0 | 4.0 | 5.0 | 21.5/25 |
| 3 | 04 | GPT-6.1 SolOpenAI | 4.5 | 3.0 | 4.0 | 4.5 | 4.5 | 20.5/25 |
| 4 | 07 | GPT-6.0 LunaOpenAI | 4.5 | 3.0 | 4.0 | 2.5 | 4.5 | 18.5/25 |
| 5 | 01 | DeepSeek FlashDeepSeek | 4.5 | 3.5 | 5.0 | 1.5 | 3.5 | 18.0/25 |
| 6 | 06 | GPT-5.6 SolOpenAI | 4.5 | 3.5 | 3.0 | 1.5 | 5.0 | 17.5/25 |
| 7 | 08 | HY3Not stated | 4.5 | 2.5 | 3.5 | 1.5 | 4.0 | 16.0/25 |
| 8 | 05 | Agnes 3.0 FlashAgnes AI | 0.5 | 1.0 | 3.5 | 2.0 | 2.5 | 9.5/25 |
Scores are editorial and based on one run per model.
Full leaderboardRubric
The 25-point test
We score the pelican, the bicycle, and — most importantly — the motion.
25points
per run
per run
Each dimension is scored 0–5. The total is out of 25.
Anatomy & contact
Feet on the pedals, wing or body connected to the handlebar, body settled on the saddle.
5 pointsBicycle structure
Two wheels, frame, fork, crank, chain, saddle and bars — the parts must form a rideable whole.
5 pointsMotion physics
Wheels turn in the direction of travel, the crank rotates, feet track the pedals, nothing slides or detaches.
5 pointsPolish & interaction
Art direction, scene detail, a complete loop, and controls that let you pause or restart.
5 pointsWhy motion
Why animation is the harder test
A still SVG asks whether a model knows what a pelican and a bicycle look like. Animation asks whether it understands how they move together: wheels turn, cranks rotate, feet follow the pedals, and the rider stays attached. Most models can draw a bicycle. Far fewer can make it move correctly.
A pretty still frame can hide a bicycle that never pedals.
Timeline
How the pelican got here
25 Oct 2024
Simon Willison asks models for an SVG of a pelican riding a bicycle. The joke becomes a benchmark.
Jun 2025
“The last six months in LLMs, illustrated by pelicans on bicycles” turns it into a public comparison.
2026
“Pelicanmaxxing” studies generate 1,000+ SVGs. Static images are still the default.
27 Oct 2026
This site extends the test to animated, interactive HTML: eight models, one prompt.
Run the same test on your AI
Copy the prompt, paste it into your model, save the result as an .html file and open it. Then check the motion, not just the still.
01
Open a model that can write and preview HTML. A chat app with a code canvas works best.
02
Paste the prompt exactly as written. Do not add constraints.
03
Ask for a single self-contained HTML file. Save it as pelican.html.
04
Open the file in a browser, watch a full loop, and score it with the checklist below.
FAQ
Questions people actually ask
The short version, before you fall down the pelican hole.
What is the pelican bicycle test?
It is an informal LLM benchmark started by Simon Willison in October 2024. He asked models to generate an SVG of a pelican riding a bicycle. The drawing looks silly, but it is a compact check of spatial reasoning, object relationships and code generation. A correct-looking result requires the model to understand how a pelican's body, a bicycle's parts and their contact points fit together.
What is the exact prompt?
The Chinese original is: “创建一个 HTML，内容是 SVG 绘制一个鹈鹕骑自行车的 2D 动画，你不需要任何测试，不要有任何限制”. In English: “Create an HTML page containing a 2D SVG animation of a pelican riding a bicycle. You don't need any tests, and don't add any restrictions.” Each model received it once, with no follow-ups.
Which model won this animated test?
Among eight runs, Doubao 2.1 Pro scored highest, 23.5 out of 25. It is the only run with inverse-kinematics legs that keep both feet on the pedals, plus a complete bicycle and a working pause/speed/bell panel. Kimi K3.1 Pro is the best-looking scene. GPT-6.1 Sol is the strongest OpenAI entry at 20.5; the two later GPT variants scored 18.5 and 17.5. DeepSeek Flash is the best still image, but its wheels and crank never move.
Why do AI models struggle with bicycles?
A bicycle is a set of rigid parts with precise contact points and a chain of motion. A pelican adds anatomy constraints on top. Models can draw plausible individual parts while failing the connections: feet hover next to pedals, wings miss the handlebar, or the wheels turn while the bike slides. A still image hides many of those failures; an animation exposes them.
Is this the same as Simon Willison's benchmark?
It is an animated extension, not a replacement. The core task comes from his October 2024 test, and the prompt lineage is credited throughout the site. Where his benchmark asks for a static SVG, this edition asks for a self-contained HTML page with a 2D SVG animation, then scores the motion.
Is this about the Little Tikes Pelican bike?
No. This is an AI benchmark. The name collision with the Little Tikes Pelican Explore & Fit Cycle, Bicicapace Pelican cargo bikes and the Box Dog Pelican Randonneur is a search-engine accident. If you are shopping for a physical bicycle, this is not the right page.
Can I reuse these outputs?
The HTML files are displayed unmodified for commentary and research. You can link to them and quote short excerpts with attribution to the model and this page. The outputs remain the property of their respective owners.
One prompt. Eight very different pelicans. Judge for yourself.
This is not a leaderboard of intelligence. It is a quick, visual stress test for spatial reasoning, code generation and — now — motion.