# Pelican Bicycle Benchmark

> One identical prompt sent to 10 consumer LLMs: an animated SVG of a pelican riding a bicycle. Every output is served unmodified and scored out of 25 on pelican anatomy, bicycle structure, motion and scene.

Tested 27–28 Oct 2026. Scores are editorial judgements against the published rubric, backed by frame-diff evidence from headless Chrome.
The test descends from Simon Willison's pelican-on-a-bicycle benchmark (25 Oct 2024).

## Results

 1. Doubao 2.1 Pro — 23.5/25 (joint first) — The only pelican that truly pedals. — https://pelican-bicycle.wwkejishe.top/models/doubao-2-1-pro/
 2. GLM 5.3 Flash — 23.5/25 (joint first) — Everything that should turn, turns — including the feet. — https://pelican-bicycle.wwkejishe.top/models/glm-5-3-flash/
 3. Kimi K3.1 Pro — 21.5/25 — The prettiest ride in the race. — https://pelican-bicycle.wwkejishe.top/models/kimi-k3-1-pro/
 4. GPT-6.1 Sol — 20.5/25 — Inverse kinematics under a coral frame. — https://pelican-bicycle.wwkejishe.top/models/gpt-6-1-sol/
 5. GPT-6.0 Luna — 18.5/25 — Lovely lake. Two frozen legs. — https://pelican-bicycle.wwkejishe.top/models/gpt-6-0-luna/
 6. DeepSeek Flash — 18.0/25 — A perfect bicycle that never pedals. — https://pelican-bicycle.wwkejishe.top/models/deepseek-flash/
 7. GPT-5.6 Sol — 17.5/25 — Beautiful coast. Runaway wheels. — https://pelican-bicycle.wwkejishe.top/models/gpt-5-6-sol/
 8. HY3 — 16.0/25 — Happy pelican. Orbiting legs. — https://pelican-bicycle.wwkejishe.top/models/hy3/
 9. DeepSeek V4.1 Flash — 15.0/25 — A sunset poster of a bicycle riding alone. — https://pelican-bicycle.wwkejishe.top/models/deepseek-v4-1-flash/
10. Agnes 3.0 Flash — 9.5/25 — The rider floats behind the bike. — https://pelican-bicycle.wwkejishe.top/models/agnes-3-0-flash/

## Pages

- [Home](https://pelican-bicycle.wwkejishe.top/): Overview, the animated scorecard and the full field.
- [Leaderboard](https://pelican-bicycle.wwkejishe.top/leaderboard/): Ranked table with per-criterion scores out of 25.
- [Compare](https://pelican-bicycle.wwkejishe.top/compare/): Synchronised side-by-side player, 2–10 lanes, deep-linkable with ?runs=slug,slug.
- [Prompt](https://pelican-bicycle.wwkejishe.top/prompt/): The exact prompt, a 10-point checklist and the reversed-prompt control run.
- [Methodology](https://pelican-bicycle.wwkejishe.top/methodology/): Scoring rubric, evidence method and the limits of this benchmark.

## 中文版本

- [首页](https://pelican-bicycle.wwkejishe.top/zh/): 总览、动画记分卡和完整参赛名单。
- [排行榜](https://pelican-bicycle.wwkejishe.top/zh/leaderboard/): 按四项标准给出 25 分制明细的排行榜。
- [并排对比](https://pelican-bicycle.wwkejishe.top/zh/compare/): 同步并排播放器，2–10 条车道，可用 ?runs=slug,slug 深链。
- [提示词](https://pelican-bicycle.wwkejishe.top/zh/prompt/): 完整提示词、10 条自查清单，以及主宾对调的反向对照运行。
- [方法](https://pelican-bicycle.wwkejishe.top/zh/methodology/): 评分标准、取证方法与本基准的局限。

## Prompts

Original (Chinese, sent verbatim):
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

English translation:
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.

Reversed control run (Chinese, sent verbatim) — subject and object swapped, Qwen 3.8 Flash:
创建一个 HTML，内容是 SVG 绘制一个自行车骑鹈鹕的 2D 动画，你不需要任何测试，不要有任何限制

## Files

- Prompt permalink: https://pelican-bicycle.wwkejishe.top/prompt.txt
- Unmodified model artifacts: https://pelican-bicycle.wwkejishe.top/artifacts/<slug>.html (slug per model, e.g. https://pelican-bicycle.wwkejishe.top/artifacts/deepseek-flash.html)
- Sitemap: https://pelican-bicycle.wwkejishe.top/sitemap.xml