Issue 2026-11-10 · Industry · 研究 · 开源
New SWE-Pair benchmark requires fixing two real bugs at once
Evaligo released SWE-Pair, 161 tasks from 69 Python projects where agents must fix two real bugs at once. DeepSeek V4.1 Flash at medium effort solved 90% at about 6 cents per task; GPT-6 Luna at max effort solved 78% at about 23 cents. More reasoning did not help and cost more.
Read original ↗