Anthropic measures learning to use the model
The Economic Index report found that given the same task, model and language, users with longer tenure get better outcomes, and that task concentration is falling on Claude.ai while rising on the API.
Why it matters
The tool and the skill of using it were separated: measured productivity depends on tenure, not only on access.
The report was published on 24 March 2026 and the sample covers 5 to 12 February 2026. The report's summary states the success-rate difference for higher-tenure users as 10 per cent; in the analysis itself the estimate is about 5 percentage points, narrowing to roughly 3 points once controls are applied. The top 10 most common tasks on Claude.ai fell from 24 to 19 per cent of conversations, while on the API they rose from 28 to 33 per cent. Tasks associated with Computer and Mathematical occupations account for 35 per cent of conversations on Claude.ai. The share of per-person usage going to the top five states fell from 30 to 24 per cent, but the pace of convergence was revised: states are now expected to reach roughly equal usage per capita in 5 to 9 years rather than 2 to 5. The average estimated value of a task fell slightly, from $49.3 to $47.9.