What Work Can Robots Do? (Anthropic, September 2026): robots can do 74% of physical tasks, are cost competitive for 0.3%

Anthropic's economics team builds an index of how well current robots can perform the physical tasks in US occupations, then combines it with LLM exposure measures. Robots can already do about 74% of physical tasks, which make up 34% of working hours, though mostly in constrained settings. Together, robots and language models touch roughly 80% of job tasks by working time, with robots covering the manual work that LLMs cannot. The binding constraint is price. Robots are cost competitive for only 0.3% of tasks today, and at the historical decline of about 3% a year in robot costs it would take around 40 years to reach 10%. Over the past 50 years, occupations with higher robot exposure saw larger declines in wages and employment, and exposure has expanded by about 2% of previously unreachable physical work each year. Driving and warehouse jobs are among the most exposed. For founders in physical AI, the report quantifies the gap between capability and economics, and points to where cost curves matter most.