ProgramBench analysis finds GPT-5.5 spends 80% of turns clarifying specifications while Claude uses iterative coding cycles · Digg