CoDA-Bench: leaderboard
RUC DataLab's CoDA-Bench (ICML 2026): 1,009 analytical questions over 199 Kaggle datasets, each posed to a code agent in a Linux sandbox holding about 980 semantically similar files across 31 data communities. The agent must find the relevant files, integrate them, write code and return the answer. The authors run each harness-and-backbone pair themselves; the score is Execution Accuracy, the share of final answers that match the reference. Higher is better.
Metric: Execution Accuracy (%). Source: coda-bench.github.io. Saturation forecast: Around February 2028. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 60.5 | #34 (GPT-5.5) |
| 2 | GPT-5.5 | 59.7 | #34 |
| 3 | Claude Sonnet 4.6 | 53.8 | #97 |
| 4 | Claude Opus 4.7 | 49.3 | #53 |
| 5 | DeepSeek V4 Pro | 49 | #113 |
| 6 | Claude Opus 4.6 | 47.8 | #69 |
| 7 | Kimi K2.6 | 43.8 | #115 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=coda-bench · How It Works · Data refreshed daily, snapshot 2026-11-07.