Active-SWE - Localization Recall: leaderboard
Active-SWE localization recall: the share of the reference patch's edit hunks that the agent's patch hits, 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 Recall (%). Source: hbinli.github.io. Saturation forecast: Around November 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 29.1 | #39 |
| 2 | GPT-5.5 | 28.8 | #33 |
| 3 | Gemini 3.1 Pro (Preview) | 26.8 | #64 |
| 4 | Kimi K2.7 Code | 18.7 | #95 |
| 5 | Qwen 3.7 Max | 18.6 | #85 |
| 6 | DeepSeek V4 Pro | 17.5 | #112 |
| 7 | GPT-5.4 | 17.2 | #83 |
| 8 | Kimi K2.6 | 17.2 | #113 |
| 9 | Claude Sonnet 4.6 | 16.5 | #94 |
| 10 | GLM-5.2 | 13.9 | #80 |
| 11 | Ring-2.6-1T | 13.3 | #176 |
| 12 | Qwen 3.5 397B A17B | 12.4 | #154 |
| 13 | Hy3 | 12.3 | #100 |
| 14 | Qwen 3.5 122B A10B | 11.2 | #184 |
| 15 | Ling-3.0-flash | 10.93 | #215 |
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-recall · How It Works · Data refreshed daily, snapshot 2026-11-05.