CoDA-Bench Hard: leaderboard
The 119-task Hard subset of RUC DataLab's CoDA-Bench, filtered to tasks that need at least two target files and a reference solution over 30 effective lines, across 15 data communities. Each sandbox averages 1,422 Kaggle files (about 2.1 GB), and the agent must discover the right data, integrate it and compute the answer. Scored by Execution Accuracy, the share of final answers matching the reference, for each harness-and-backbone pair the authors ran. Higher is better.
Metric: Execution Accuracy (%). Source: coda-bench.github.io. Saturation forecast: Around 2030. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 45.4 | #69 |
| 2 | GPT-5.5 | 44.5 | #34 |
| 3 | Claude Sonnet 4.6 | 42.9 | #97 |
| 4 | GPT-5.5 (xHigh) | 42.9 | #34 (GPT-5.5) |
| 5 | Claude Opus 4.7 | 38.7 | #53 |
| 6 | Kimi K2.6 | 37 | #115 |
| 7 | DeepSeek V4 Pro | 36.1 | #113 |
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-hard · How It Works · Data refreshed daily, snapshot 2026-11-07.