Robots can do 74% of US physical tasks, but are cheaper than people for only 0.3%

Anthropic published a study on 30 September 2026 that puts a number on a fear and then undercuts it. Its researchers estimate that robots can already perform 74% of the physical tasks in the US economy, which account for 34% of working hours, but that robots are cheaper than people for only 0.3% of tasks. The gap between those figures is the story, and it matters more than either headline.

What Anthropic measured

The paper, titled "Can we predict the jobs robots will do?", builds a robot exposure index. According to Anthropic's write-up, researchers used Claude to score 7,594 physical tasks drawn from O*NET, the US Labor Department's occupational database, against a four-tier scale from E0 to E3. The tiers depend on how controlled the environment has to be: a factory floor or warehouse is easy, a city road or a cluttered home is hard.

Robots are defined narrowly as autonomous physical machines that sense and act. An automated car wash counts. A surgeon's directly controlled instrument does not. Tasks were checked against the existence of real robots, not just plausible ones.

The numbers that matter

• 74% of physical tasks can be done by robots that exist today, mostly in controlled settings, covering 34% of working hours.
• Only 0.3% of tasks are cost competitive with human labour now, per Anthropic's analysis.
• At a historical price decline of 3% a year, it would take around 40 years for that to reach 10% of tasks.
• Combining robot and large language model exposure, about 80% of employment, by working time, is exposed to at least one of the two technologies.
• Only about 1% of tasks involve robots working in unstructured settings such as city roads, as reported by Mixed.

The most useful sentence in the paper is its own caveat.

> "While robots can do most physical work tasks today, they are much more expensive than human labor." (Anthropic, "Can we predict the jobs robots will do?", 30 September 2026)

Why cost is the real variable

Exposure is a ceiling, not a forecast. A task being technically possible tells you nothing about whether anyone will pay for the machine. Anthropic names three further brakes beyond price: capability gaps (the paper's example is untangling wire), human preferences, and regulation. That is a more careful position than the one that usually travels in headlines about automation.

It also means the 40 year figure should be read as arithmetic, not prophecy. It assumes a steady 3% annual decline in robot costs. If costs fall faster the timeline compresses sharply, and if they stall it stretches. Small changes in that one assumption move the answer by decades. We would not quote it as a date.

Who is exposed first

Anthropic says driving and warehouse work are the most exposed to current robots, with taxi drivers and warehouse packers named as facing change sooner than nurses and mechanics. Nursing and general repair are not, because present day robots can do little of that work even in highly controlled conditions.

The people in exposed jobs are, per the paper, more likely to be male, less educated and lower paid. That is the opposite of the profile most often attached to the large language model wave, which tends to hit office workers first. The two together leave only about one fifth of employment outside the reach of either technology.

There is also a backward looking result. Looking at the period since 1977, Anthropic reports that jobs more exposed to robots saw bigger falls in wages and employment than others. That is a correlation in US data, and the paper should not be read as proving that the next wave will follow the same pattern.

What it means for policy

The cost gap is also where the political argument will be fought.

The study lands in a live US argument about who pays when jobs disappear. A senator's bill would tax AI companies to fund support for displaced workers, which we reported on 28 September. Anthropic's data cuts both ways in that debate. It shows large technical exposure, which supports the urgency argument, and a tiny cost competitive share, which supports the case that there is time to prepare.

The UK reading

This is a US study built on US occupational data. We found no equivalent British analysis in the sources we opened, and the 74%, 34% and 0.3% figures should not be applied to the UK labour market as they stand. What carries over is the method: ask what a machine can do, then ask whether anyone can afford it. British employers and unions asking about automation should insist on both questions.

What to watch

Anthropic is also a company with commercial reasons to describe how capable AI is, and the exposure scores were produced by its own model. That is a method worth probing: the scores are judgements by Claude, checked against web searches, not a measurement of deployed robots. An independent replication using a different model or human raters would tell us how robust the 74% is.

Our take

The study is careful, and its author is open about the limits.

The number to remember is 0.3%, not 74%. A study that finds near universal technical exposure and almost no economic displacement today is describing a long runway, not a cliff. Plan for it, argue about who pays for it, and treat anyone who turns the 74% into a jobs forecast with suspicion.