Active-SWE: leaderboard
Active-SWE asks coding agents to find and fix bugs without issue reports, on 400 tasks in 8 languages (arXiv 2608.04682v1). Resolved is the percentage of recorded-bug tasks whose fix passes all reference tests, 0-100, higher is better. Every model ran in the Claude Code harness; from the project site's data/site-data.json (full 400-task main set).
Metric: Resolved (%). Source: hbinli.github.io. Saturation forecast: Around July 2028. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 20 | #39 |
| 2 | GPT-5.5 | 18.5 | #33 |
| 3 | Gemini 3.1 Pro (Preview) | 15.8 | #64 |
| 4 | Kimi K2.7 Code | 14.8 | #95 |
| 5 | Kimi K2.6 | 14 | #113 |
| 6 | GLM-5.2 | 13 | #80 |
| 7 | Qwen 3.7 Max | 11.8 | #85 |
| 8 | Ring-2.6-1T | 11 | #176 |
| 9 | Hy3 | 9.8 | #100 |
| 10 | DeepSeek V4 Pro | 9.5 | #112 |
| 11 | GPT-5.4 | 8.8 | #83 |
| 12 | Claude Sonnet 4.6 | 8.3 | #94 |
| 13 | MiniMax-M2.7 | 7.8 | #270 |
| 14 | Ling-3.0-flash | 7.75 | #215 |
| 15 | Qwen 3.5 397B A17B | 6.8 | #154 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=active-swe · How It Works · Data refreshed daily, snapshot 2026-11-05.