Anthropic’s economics team analyzed roughly 400,000 Claude Code sessions from about 235,000 users between October 2025 and April 2026, using a privacy-preserving method. The central finding is that how well someone understands the problem they are solving predicts success more reliably than whether they were formally trained to write code. Across the ten largest occupations in the data, every one landed within about seven points of software engineers on code-producing sessions.
Task-specific expertise, by contrast, showed large returns. Expert-rated users reached verified success in roughly 28-33 percent of sessions versus about 15 percent for novices. Experts also drove the agent harder, triggering about 12 actions per prompt compared with 5 for novices, and abandoned troubled sessions far less often (5-7 percent versus 19 percent). In a typical session people made most of the planning decisions while Claude made most of the execution decisions, and more domain expertise meant Claude did more work per instruction.
The composition of work also shifted over the seven months: debugging fell from 33 to 19 percent of activity while software operation and data analysis nearly doubled from 10 to 20 percent, and average task value rose 27 percent. The authors read this as evidence that agentic coding tools reward domain expertise rather than replacing it.