On September 30, 2026, Anthropic economists Russell Legate-Yang and Maxim Massenkoff published a robot exposure index asking which jobs robots can actually do today. They used Claude to rate 7,594 physical job tasks from the US O*NET occupational database on a four-level rubric: robots cannot do the task (E0), can do it only in purpose-built environments (E1), in structured human facilities (E2), or in unstructured environments (E3). They then joined the ratings to BLS employment and wage data and to Census survey data.
The study separates what is technically possible from what is economically viable:
- Robots "can perform three-quarters of physical tasks in the US, making up 34% of working hours, but mostly in limited settings."
- But "robots are cost-competitive for just 0.3% of job tasks." A task counts as cost-competitive when the robot's cost is below the labour cost of that slice of the job. "If robot price declines follow past trends, it will take 40 years for that share to reach 10%."
- Across robots and language models together, "about 80% of job tasks by working time are exposed to either robots or LLMs."
- Exposure falls on lower-paid workers: highly exposed workers "earn around $30 less per hour," according to the post, and face an unemployment rate more than twice as high. Taxi drivers top the index at 2.2 out of 3. Packers and packagers are the largest occupation exposed to cost-competitive robots, and their employment "has fallen 22% since 2015."
- The authors back-test the method: from 1977 onward, jobs more exposed to the robots of the day saw wage and employment declines in later decades.
Why it matters: most AI labour-market research so far has measured exposure to language models, which mainly affects office work. This study points at a different and lower-paid workforce, and its main finding is a gap: robots can technically do most physical work, but the economics only work for a tiny fraction of it today. Caveats: the capability ratings come from a language model rather than field trials, cost estimates are approximations, and the 40-year projection assumes uniform price declines. The authors themselves note that AI-driven robots could "leapfrog" these estimates. Treat robot cost curves, not capability demos, as the leading indicator worth watching.
Anthropic finds robots can technically handle three-quarters of US physical tasks but are cost-competitive for just 0.3%, so robot cost, not capability, is the binding constraint, and the exposure falls on lower-paid workers.