Robots can already do three quarters of physical work, but they pay off in just 0.3% of cases: Anthropic's analysis

On 30 September 2026, Anthropic researchers published a study titled "What work can robots do?", which sets out to map a robot exposure index for the US economy. Working through 7,594 physical tasks taken from the US Department of Labor's O*NET database, the researchers used the Claude language model to compare what each task requires with documented uses of robots found on the web. The ranking has four levels of environmental complexity, from tasks a robot cannot perform at all up to operating in unstructured settings such as city streets.

The results show a sharp gap between technical feasibility and financial sense. According to the analysis, robots could already perform 74% of physical tasks in the United States, equivalent to 34% of all hours worked. Almost always, though, that holds in environments built specifically for machines, like an assembly line, or at least orderly ones, like a logistics warehouse. It is far rarer in unstructured settings like a city street. And the economics stop at just 0.3% of tasks. The authors count a task as cost-effective when a robot's annual cost (hardware, installation, maintenance and operation, with fixed costs spread over ten years and an 8% cost of capital) is lower than the cost of the worker, meaning wages plus benefits, for the time spent on those tasks. The estimates of robot costs and human productivity come from Claude. If machine prices keep falling at the historical rate the authors estimate, about 3% a year, it would take forty years for the cost-effective share to reach 10%.

Combine the impact of robots with that of language models and overall exposure rises to about 80% of hours worked. Taxi drivers are among the most exposed groups, while healthcare and repair work remain the least affected. The authors note that the most exposed workers are mostly men with lower levels of education and pay. They acknowledge several limitations: O*NET descriptions are brief and need interpretation, cost estimates are rough, the same price decline is applied to every task, and regulatory uncertainty and human preferences are left out of the model. I would add that capabilities and costs are estimated by Claude, Anthropic's own model, and have not been checked in the field. The study has not yet been independently peer-reviewed, and the forty-year projection ignores possible technological leaps.

The distance between being able to do something and it being worth doing reminds us that automation is not a linear technological destiny but an accounting equation. As long as hardware remains such a heavy investment, human flexibility will, paradoxically, stay the cheapest option on the market.

— Olya