Vaibhav (VB) Srivastav (@reach_vb)

GPT-5.6 Luna Max scores 13.3 points above Sonnet 5 Max on DeepSWE v1.1 while Sonnet costs 44x as much. DeepSWE tests coding agents on 113 original, long-horizon engineering tasks. Luna posts 67.2% at $0.61/task, 1.7 points above Gemini 3.7 Flash Medium at 70% lower cost.