BulkPR-Bench: leaderboard
BulkPR-Bench (2026) gives a coding agent a queue of 581 newly authored, interacting pull requests on frozen snapshots of 18 real repositories, released in batches, and asks it to merge a large safe subset in executable order and reject the rest. The Relational Delivery Score credits safe delivery and correct rejection within each execution-validated relation group, in percent; higher is better. Every model runs in the same Claude Code scaffold.
Metric: Relational Delivery Score (%; 581 newly authored candidate pull requests on frozen snapshots of 18 real repositories, released in batches of K=32 under the buffered protocol (up to 4 candidates deferred for up to 16 batches); credit for safe delivery in executable order and correct rejection within each gold relation group of the realized merge trace, repository-weighted, three runs per repository; every model runs in the same Claude Code v2.1.138 scaffold with a 150-turn budget). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 66.6 | #56 |
| 2 | GPT-5.4 | 62 | #123 |
| 3 | GLM-5.2 | 57.9 | #138 |
| 4 | Kimi K3 | 49.4 | #54 |
| 5 | Qwen 3.7 Max | 47.2 | #145 |
| 6 | DeepSeek V4 Pro | 40.9 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bulkpr-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.