Active-SWE - Test Validity: leaderboard
Active-SWE test validity on potential bugs: the percentage of tasks where every test the agent generates fails before and passes after the fix (arXiv 2608.04682v1). 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: Test Validity (%). Source: hbinli.github.io. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 87.5 | #33 |
| 2 | Claude Opus 4.8 | 81 | #39 |
| 3 | Qwen 3.7 Max | 81 | #85 |
| 4 | GLM-5.2 | 79 | #80 |
| 5 | Claude Sonnet 4.6 | 78.3 | #94 |
| 6 | Gemini 3.1 Pro (Preview) | 77.8 | #64 |
| 7 | Hy3 | 77.3 | #100 |
| 8 | Kimi K2.6 | 67.3 | #113 |
| 9 | Ring-2.6-1T | 65.5 | #176 |
| 10 | Kimi K2.7 Code | 65 | #95 |
| 11 | DeepSeek V4 Pro | 64.3 | #112 |
| 12 | Qwen 3.5 397B A17B | 56.8 | #154 |
| 13 | GPT-5.4 | 52.3 | #83 |
| 14 | Ling-3.0-flash | 49 | #215 |
| 15 | Qwen 3.5 122B A10B | 46.8 | #184 |
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-test-validity · How It Works · Data refreshed daily, snapshot 2026-11-05.