ProcCtrlBench: leaderboard
Execution-process quality of LLM coding agents (He et al., 2026): trajectories of each agent and model on 200 tasks from SWE-bench Verified, TerminalBench and AndroidBench are checked by calibrated detectors for 11 process-defect types and for control preservation (interpretable, interruptible, correctable, reversible execution). Summary score PB from 0 to 1; higher is better.
Metric: ProcCtrlBench score PB (0-1; calibrated process quality over 11 execution-defect types combined with control preservation, 200 tasks: 100 SWE-bench Verified, 60 TerminalBench, 40 AndroidBench). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.74 | #158 |
| 2 | Claude Sonnet 4.5 | 0.72 | #244 |
| 3 | GPT-5.4 | 0.7 | #121 |
| 4 | Claude Sonnet 4 | 0.69 | #397 |
| 5 | GPT-5 | 0.69 | #163 |
| 6 | Qwen 3 Coder Plus | 0.67 | #712 |
| 7 | Qwen 3.5 Plus | 0.66 | #227 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=procctrlbench · How It Works · Data refreshed daily, snapshot 2026-11-12.