Hacker News
Anthropic's 400K-Session Study: Domain Expertise, Not Coding Background, Decides Whether Claude Code Succeeds
Anthropic's new 'Agentic coding and persistent returns to expertise' research analyzed ~400,000 Claude Code sessions (235,000 people, Oct 2025–Apr 2026) and found expert users hit 33% verified success vs 28% intermediate and 15% novice — more than 2x. Experts trigger ~12 Claude actions and ~3,200 words per prompt vs 5 actions / ~600 words for novices, and nearly every occupation succeeds at coding tasks at roughly the same rate as software engineers. The actionable read for builders: humans still own planning ('what'), the model owns execution ('how'), and the scarce input is understanding the problem — not typing code.
Source
↳ Follow the thread