LongDA - Match - 100 Steps: leaderboard
Evaluates agents performing data analysis over long survey documentation: 505 queries from 30 publications, using the LongTA scaffold with a 100-step budget. Match rate measures numeric answers within the paper's 5% tolerance. Coverage is output-format compliance and resource columns are diagnostics, not this result. Explicit High and Thinking model labels are retained. This series reports LongDA - Match - 100 Steps using Numeric Match (%).
Metric: Numeric Match (%). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 69.16 | #105 |
| 2 | GPT-5 (High) | 68.91 | #105 (GPT-5) |
| 3 | DeepSeek V3.2 | 53 | #210 |
| 4 | DeepSeek V3.2 (Thinking) | 47.82 | #210 (DeepSeek V3.2) |
| 5 | GPT-5 Mini | 40.81 | #193 |
| 6 | Kimi K2 0905 | 28.27 | #292 |
| 7 | Qwen 3 235B A22B Instruct | 27.17 | #336 |
| 8 | Qwen 3 Coder 480B A35B Instruct | 23.44 | #314 |
| 9 | Qwen 3 Next 80B A3B Instruct | 21.85 | #381 |
| 10 | GLM-4.7 | 19.18 | #202 |
| 11 | GPT-OSS-120B | 12.15 | #346 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=longda-match-100-steps · How It Works · Data refreshed daily, snapshot 2026-11-05.