Active-SWE - Localization Precision: leaderboard
Active-SWE localization precision: the share of the agent patch's edit hunks that hit the reference patch, over the recorded-bug tasks (arXiv 2608.04682v1). Percentage 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: Localization Precision (%). Source: hbinli.github.io. Saturation forecast: Around June 2028. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 35 | #39 |
| 2 | GPT-5.5 | 24.1 | #33 |
| 3 | Qwen 3.7 Max | 19.4 | #85 |
| 4 | DeepSeek V4 Pro | 18.9 | #112 |
| 5 | Gemini 3.1 Pro (Preview) | 18.8 | #64 |
| 6 | Claude Sonnet 4.6 | 18 | #94 |
| 7 | GPT-5.4 | 17.6 | #83 |
| 8 | GLM-5.2 | 17 | #80 |
| 9 | Kimi K2.7 Code | 16.1 | #95 |
| 10 | Hy3 | 15.8 | #100 |
| 11 | Ring-2.6-1T | 14.9 | #176 |
| 12 | Kimi K2.6 | 14.7 | #113 |
| 13 | Qwen 3.5 397B A17B | 13.9 | #154 |
| 14 | Qwen 3.5 122B A10B | 13.8 | #184 |
| 15 | MiniMax-M2.7 | 13.4 | #270 |
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-localization-precision · How It Works · Data refreshed daily, snapshot 2026-11-05.