SWE-Test (Claude Code): leaderboard
LLM vulnerability discovery via input prediction in real software targets, driven by the Claude Code scaffold in feedback-enabled fixed-target mode; pass rate over valid trials. Other scaffolds in the paper cover one model each and are not published. Higher is better. Source: arxiv.org/abs/2609.06229.
Metric: Feedback-enabled Pass Rate (%). Source: arxiv.org. Saturation forecast: Around July 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max | 55 | #34 |
| 2 | GLM-5.2 | 41.7 | #80 |
| 3 | DeepSeek V4 Pro | 35 | #112 |
| 4 | Kimi K2.6 | 35 | #113 |
| 5 | Kimi K3 | 35 | #35 |
| 6 | Qwen 3.7 Max | 30 | #85 |
| 7 | GLM-5.1 | 23.3 | #109 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swe-test-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-05.