DSAEval: leaderboard
Data-science agent benchmark of 641 open-ended problems over 285 real-world datasets, structured and unstructured (text, images), solved by an agent that writes and runs code over several steps. Each run is scored 0-10 as 0.3 x reasoning + 0.3 x code + 0.4 x final result, averaged over two LLM judges (Claude Haiku 4.5 and GPT-5.1); higher is better. Hong Kong Polytechnic University, EMNLP 2026.
Metric: Total score (0-10; mean of two LLM judges). Source: ama-cmfai.github.io. Saturation forecast: Around December 2026. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 8.16 | #244 |
| 2 | MiMo-V2-Pro | 7.91 | #324 |
| 3 | GPT-5.2 | 7.71 | #172 |
| 4 | MiniMax-M2.7 | 7.7 | #519 |
| 5 | MiMo-V2-Flash | 7.64 | #674 |
| 6 | MiniMax-M2 | 7.64 | #669 |
| 7 | Gemini 3 Pro | 7.31 | #135 |
| 8 | Grok 4.1 Fast | 7.25 | #426 |
| 9 | GPT-5 Nano | 7.07 | #888 |
| 10 | DeepSeek V3.2 | 7.03 | #403 |
| 11 | GLM-4.6V | 6.87 | #698 |
| 12 | Qwen 3 VL 30B A3B (Thinking) | 5.32 | #710 |
| 13 | Ministral 3 14B | 5.18 | #1852 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dsaeval · How It Works · Data refreshed daily, snapshot 2026-11-12.