Anthropic researchers analyzed roughly 400,000 interactive Claude Code sessions from approximately 235,000 users between October 2025 and April 2026, finding that verified task success rates across non-software occupations (26%) trail software engineers (30%) by only four to five percentage points, while domain expertise within a session predicts success far more strongly: novice-rated sessions reach verified success 15% of the time versus 28–33% for intermediate-to-expert sessions. The study also documents a structural shift in usage over seven months: debugging fell from 33% to 19% of sessions, end-to-end agentic work (deployment, data analysis, document writing) roughly doubled, and the estimated freelance-equivalent value of the average session rose approximately 27%. For financial institutions, this is an early empirical signal that agentic coding tools may be broadening the population of workers who can complete technical tasks, with implications for how AI-driven productivity gains distribute across knowledge-work roles.
347 other tracked developments touching the same claims.