Pelican Benchmark
What AI models draw when asked for an SVG of a pelican riding a bicycle.
The pelican-riding-a-bicycle test was proposed and popularized by Simon Willison. This site collects community submissions; it is not affiliated with him.
The prompt
Generate an SVG of a pelican riding a bicyclePaste it into any language model, then submit the SVG it returns — good, bad or odd-looking.
Looking for moving pelicans? The Animated HTML track asks for a self-contained HTML animation.
Two takes on the same prompt
These submissions record the same prompt but different thinking levels. Open the examples page for reported settings and notes.
About the pelican test
The test asks a language model to write SVG code, not to generate a picture with an image model. A pelican and a bicycle are awkward to combine: the bird has to sit on the frame, reach the pedals and hold the handlebars. So the drawing shows, at a glance, how a model turned a short instruction into a coherent structure. It is a fun, informal check, not a scientific measurement.
How to read these results
Each card is one submission: one output from one model, as reported by the person who submitted it.
- The model name, thinking level and prompt are what the contributor entered. We do not verify them.
- The generation mode says whether the result came from a single reply, several rounds of feedback, or human edits to the code.
- The same model can draw very differently from one run to the next, and contributors may have used different settings or picked their favourite run. One result does not stand for the model.
There are no scores or rankings here. See how the collection is put together on the How this collection works page.
Submit your own result
- Copy the prompt above and send it to a model.
- Save the SVG code it returns. Clumsy or odd-looking pelicans are welcome too.
- Upload it with the model, thinking level and the prompt you actually used. No account needed.
FAQ
What is the pelican test?
A one-line prompt, “Generate an SVG of a pelican riding a bicycle”, used to see how language models handle drawing with code. Simon Willison proposed it and uses it to try out new models.
How do I take part?
Run the prompt on any model and submit the file it returns on the submit page. You can use a different prompt, but you must enter the prompt you actually used.
Are model names verified?
No. Model and generation details are provided by the contributor. We check that a submission is a readable pelican-and-bicycle drawing and is not abusive, but we cannot confirm which model made it.
Is there a score or ranking?
No. Results are listed newest first, and you can browse them by model. See How this collection works.