DSGym - DABStep Hard: leaderboard
Hard split of DABStep's financial data-analysis questions, which need multi-step reasoning across several data files, run by DSGym's default CodeAct agent (Nie et al., 2026) in an isolated Jupyter container at temperature 0 with all tools disabled. Exact-match accuracy in percent; higher is better. This harness differs from the agent submissions on the official DABStep leaderboard.
Metric: Exact-Match Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 37.04 | #244 |
| 2 | Claude Sonnet 4 | 31.75 | #397 |
| 3 | Kimi K2 0905 | 28.84 | #555 |
| 4 | GPT-5 (Medium) | 28.31 | #163 (GPT-5) |
| 5 | DeepSeek V3.1 | 21.96 | #500 |
| 6 | GPT-5.1 (High) | 13.23 | #225 (GPT-5.1) |
| 7 | GPT-5.1 (Non-reasoning) | 11.9 | #225 (GPT-5.1) |
| 8 | GPT-OSS-120B | 7.94 | #699 |
| 9 | GPT-4o (2024-08-06) | 7.41 | #731 |
| 10 | Qwen 3 4B Instruct | 2.9 | #1660 |
| 11 | Qwen 2.5 7B Instruct | 2.38 | #2583 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dsgym-dabstep-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.