9h agoProgramBench analysis finds GPT-5.5 spends 80% of turns clarifying specifications while Claude uses iterative coding cyclesGPT-5.5 delivers near-final code with minimal verification afterward.