BridgeBench V3 Arena Debugging
Code · 2026-07-10
The BridgeBench V3 Debugging arena asks models to isolate the true cause of a failing system despite planted decoys and symptom-hiding fixes. Its 18 tasks span six debugging clusters and use blind pairwise judging.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 1156 |
| GLM-5.2 | 1002 |
| Grok 4.5 | 992 |
| GPT-5.6 Sol | 848 |