Code-QA-Bench: leaderboard
Repository-level code understanding benchmark (Code-QA-Bench; Zhang et al., Baidu, 2026): 528 questions about 10 Python repositories from SWE-bench, each derived from a gold answer verified against the source code, are answered by an agent with file-reading and code-search tools in a copy of the repository stripped of docstrings, comments and documentation. A GPT-5.4 judge scores accuracy, completeness and specificity (0-5 each) against the gold answer and rubric, and the score is their mean scaled to 0-1; higher is better.
Metric: Judge score (0-1; mean of accuracy, completeness and specificity; 528 code-derivable questions, documentation-stripped repository, GPT-5.4 judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 0.89 | #181 |
| 2 | Claude Opus 4.6 | 0.89 | #100 |
| 3 | Kimi K2.6 | 0.87 | #173 |
| 4 | Gemini 3.1 Pro (Preview) | 0.77 | #91 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=code-qa-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.