CodeGolf Bench - Python Best Percentile: leaderboard
CodeGolf Bench Python Best Percentile places the shortest correct Python solution an LLM produces for each code.golf problem within the distribution of human solutions, from 0 (failing or worse than every human) to 100 (shorter than every human), and averages over problems (CodeGolf Bench). Higher is better.
Metric: Best percentile (0-100): for each problem, the best of the sampled Python solutions placed in the distribution of human code.golf solutions by character count (0 = worse than every human or failing, 100 = shorter than every human), averaged over problems; 115 code.golf problems (holes), solutions checked against the platform's hidden tests through its API; golfing prompt, temperature 0.2, top-p 0.95, up to 32,768 output tokens, 10 samples per problem; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 03-25) | 68.04 | #241 |
| 2 | DeepSeek R1 | 67.09 | #482 |
| 3 | Llama 4 Maverick | 55.31 | #1110 |
| 4 | DeepSeek V3 (0324) | 51.6 | #715 |
| 5 | Gemini 2.5 Flash (Preview 04-17) | 46.42 | #453 |
| 6 | DeepSeek R1 Distill Llama 70B | 42.33 | #1669 |
| 7 | Qwen 3 235B A22B | 41.64 | #628 |
| 8 | Llama 4 Scout | 36.69 | #1818 |
| 9 | Gemma 3 27B (IT) | 25.23 | #1324 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=codegolf-bench-python-best-percentile · How It Works · Data refreshed daily, snapshot 2026-11-21.