Anthropic reports early estimates of Claude's productivity gains
Analysing 100,000 Claude.ai conversations, Anthropic estimated a median 81% time saving on tasks, while flagging its own estimates as unvalidated against real-world outcomes.
- Money & business
- Colour
Anthropic published an analysis of roughly 100,000 anonymised Claude.ai conversations, in which the model itself estimated how long a task would take a human professional to complete versus how long the same task took with Claude’s help. Mapping those task-level estimates onto occupational categories and wage data, Anthropic reported a median time saving of around 81% and projected the effect, extrapolated across the labour market, could roughly double recent US labour-productivity growth over the coming decade.
Anthropic was explicit about the method’s limits: Claude’s own time estimates had not been validated against how long tasks actually take people, the comparison rested on assumptions about hiring a human professional to do the same work, and the analysis could not capture time spent by humans on a task outside the conversation itself, nor any effects from organisations restructuring work around AI tools. The sample was also limited to Claude.ai usage, which the company acknowledged likely skewed toward tasks Claude handles well.
The report was one of several attempts by frontier labs in 2025 to quantify AI’s economic effect using their own usage data rather than external surveys, and it drew both interest and scepticism for using a model’s self-reported estimate of the counterfactual as its central measurement.