Skip to content
Computing Life · Share · Yage

Anthropic measured how AI coding habits shifted across 400,000 Claude Code sessions

Anthropic analyzed 400,000 real Claude Code sessions from Oct 2025 to Apr 2026. Debugging sessions dropped from 33% to 19%, while ops and writing/data analysis each roughly doubled. About 56% of sessions involve direct coding; over 40% don't center on writing code. Humans handle ~70% of planning decisions, agents ~80% of execution decisions. Novice verification success is ~15%, intermediate+ jumps to 28%–33%, but in hard sessions novices succeed only 4% vs. 15% for experts. Three practice directions: add acceptance criteria before prompting, locate the deviation point when stuck instead of restarting, and break tasks into multi-step workflows. Data covers interactive sessions only, no CI pipeline calls. Success signals are CI pass, green tests, or user confirmation—no long-term maintenance cost measured.

Why it matters: Anthropic analyzed 400,000 real Claude Code sessions: bug fixing fell from 33% to 19%, while ops and writing/data analysis each doubled, with numbers on how humans and the model split decisions.

Read the original ↗Export Markdown