DARE-bench (Data Science) - Classification (Instruction Following): leaderboard
Metric: Strict accuracy (%) on the 74 classification instruction-following tasks: the agent must reproduce a reference workflow exactly and scores 1 only when its predictions equal the reference solution's; DARE-bench test tasks derived from recently updated Kaggle datasets; the model works as a data-science agent with a sandboxed Python execution tool (5 interaction turns, 200 s per execution, greedy decoding), mean of three repeats; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 69.81 | #91 |
| 2 | O4 Mini | 67.56 | #172 |
| 3 | Claude 3.7 Sonnet | 61.48 | #241 |
| 4 | GPT-4.1 | 55.82 | #240 |
| 5 | GPT-4o | 32.88 | #333 |
| 6 | Qwen 3 32B | 17.11 | #424 |
| 7 | Claude Sonnet 4 | 16.21 | #194 |
| 8 | Qwen 3 4B | 3.6 | #823 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=dare-bench-data-science-classification-instruction-following · How It Works · Data refreshed daily, snapshot 2026-10-11.