Details
- Anthropic unveils new economic research introducing a framework to track how its coding assistant Claude Code is used as it scales.
- The study uses Anthropic's privacy-preserving analysis tool to examine 400,000 Claude Code sessions between October 2025 and April 2026.
- Sessions are classified by main goal: over half involve writing or repairing code, while nearly 20% focus on operating software.
- Researchers estimate task value by comparing each session's work to equivalent tasks on freelance marketplaces, finding the average session's monetary value rose 27% over the study period.
- Success rates are compared across occupations using a stringent metric requiring verifiable completion (such as committed code), with every field within 7 percentage points of software engineers.
- Domain experts, identified via their questions and vocabulary, achieve higher success rates, but the performance gap between intermediate and expert users is modest.
- Findings suggest that solid domain proficiency, rather than deep expertise, is generally sufficient to code effectively with Claude Code.
- Anthropic plans to incorporate these new measures into the Anthropic Economic Index to monitor shifts in the nature and value of work over time.
- The report positions Claude Code as both a coding tool and an economic data source for understanding how AI-enabled coding tasks evolve across sectors and skill levels.
Impact
By quantifying how Claude Code is used and how task value changes over time, Anthropic adds a structured lens for evaluating AI’s impact on software work and adjacent occupations. Integrating these metrics into the Anthropic Economic Index could influence how policymakers, enterprises, and investors assess productivity gains, skill requirements, and labor-market shifts from AI-assisted coding.
