Anthropic finds robots can do most physical tasks but rarely cheaply enough to use

Anthropic published economic research estimating how much of America’s physical work robots can already perform, and found a wide gap between physical capability and economic viability. Using Claude to score 7,594 physical job tasks drawn from the US Department of Labor’s O*NET database against a four-tier scale of how controlled an environment a robot needs to do them, the study — by Anthropic researchers Russell Legate-Yang and Maxim Massenkoff — found that robots can already perform some version of 74% of physical tasks, accounting for 34% of all US working hours.

Cost, not capability, was the binding constraint: robots were cost-competitive with human labour for only 0.3% of job tasks today, and extrapolating from historical price declines of roughly 3% a year, the researchers estimated it would take about 40 years to reach just 10% cost-competitiveness. Workers in the most robot-exposed occupations were more likely to be male, far less likely to hold a bachelor’s degree, earned around $30 an hour less, and faced more than double the unemployment rate of unexposed workers; a backtest against 50 years of wage and employment data found that more exposed occupations had already fared worse on both measures. Combined with tasks exposed to language models, Anthropic said about 80% of all US working hours were exposed to one automating technology or the other.

The report extended the methodology Anthropic had used since launching its Economic Index in February 2025 and refined in later instalments tracking white-collar task exposure, applying the same O*NET framework to the physical side of the automation debate for the first time.