The Robots Can Do the Job — Just Not at the Price
Anthropic published a new index on September 30, 2026 measuring how exposed American jobs are to robots. The headline finding is a striking gap: robots can already perform 74 percent of physical tasks in the US — equivalent to 34 percent of working hours — yet they are cost-competitive for just 0.3 percent of job tasks. The conclusion of the company's own research is that this gap between capability and economics, not raw technical ability, is what predicts which physical occupations change first. But all the key figures come from Anthropic's own analysis, in which the company's own AI model, Claude, scores the task descriptions (source: 39cd6c19) — and the index measures exposure, not layoff probability.
What's New
The index extends Anthropic's earlier work on language models' exposure in the labor market, published in March 2026. That work concerned how much of the tasks in various occupations can be covered by large language models (LLMs). The new index turns to physical work and asks: how much of it can robots do today?
Anthropic's own answer is unambiguous on the capability side: the company writes in the research post that robots can already perform 74 percent of physical tasks in the US, which account for 34 percent of working hours (Anthropic, 39cd6c19). Measured in working time, then, roughly a third of the physical content of American working life is technically within reach, according to the company.
The economics look entirely different. Only 0.3 percent of job tasks are currently cost-competitive for robots. And if robot prices continue falling in line with historical trends, Anthropic estimates it will take 40 years before that share reaches 10 percent (39cd6c19).
Taken together, the company concludes that around 80 percent of job tasks, measured in working time, are exposed to either robots or language models. "Robots do work where LLMs cannot," the post states (39cd6c19). What remains outside exposure, according to the company, is work that is strongly interpersonal, or that requires physical skills today's robots lack.
How the Index Is Built
Methodologically, the study is built on O*NET, an American database linking roughly 900 occupations to descriptions of around 19,000 job tasks. Anthropic first identifies a set of physical tasks that cannot be automated without robots. The company then has Claude score the task descriptions according to a rubric measuring the demands for physical, cognitive, and interpersonal work (39cd6c19).
This is where one of the study's most important methodological caveats lies: it is a commercial AI model — developed by Anthropic itself — that performs the actual assessment work. The listing of occupations and tasks from O*NET is publicly available, but the scoring depends on how Claude interprets the descriptions and the rubric. The sources do not provide full insight into every detail of the methodology, including exactly how cost-competitiveness is calculated. The figures should therefore be read as Anthropic's own estimates, not independently verified findings — the study is self-published, and independent replication or peer review is not documented in the available source material.
The Gap: Can Do, Rarely Pays Off
The analytical core of the study is the distinction between two questions that are often conflated: can the robot do the job, and does it pay to use it?
On the first question, capability is, according to the study, surprisingly high. On the second, it is nearly absent. The difference between 74 and 0.3 percent illustrates that, by Anthropic's figures, technical possibility and economic rationality follow entirely different timescales: a robot may be capable of performing a task yet still cost more than human labor.
Anthropic's estimate of 40 years to reach 10 percent cost-competitiveness rests on historical price trends continuing. It is a conditional projection, not a guaranteed path — and the sources do not describe the details behind the calculation model, so it is difficult to assess on its own.
At the same time, the study shows that capability is not standing still. According to Anthropic's figures, robots have each year become able to perform roughly 2 percent of the physical work they previously could not master (39cd6c19). It is a slow but steady expansion of what machines can do.
The Index Actually Tracks Reality, the Backtest Shows
The strongest argument that the index captures something real is a retrospective analysis spanning 50 years. According to Anthropic, occupations with high robot exposure saw larger declines in wages and employment than other occupations over this period. Meanwhile, exposure has grown steadily, in step with the roughly 2 percent annual capacity growth (39cd6c19).
This is an internal validation of the index: it appears to distinguish occupations that have historically been squeezed from those that have not. But it is worth noting that the backtest also comes from the same research effort — it has not been independently replicated in the available material.
Who Is Affected — and Who Is Not
The index points clearly to which occupational groups rank highest first. Driving and warehouse work are, according to Anthropic, highly exposed to robots that already exist. Nursing and general repair work, by contrast, are not — today's robots can, according to the company, do little of this work, even in highly controlled environments (39cd6c19).
Demographically, the picture is clear, according to the study: workers exposed to robots are more likely to be men, with lower education and lower wages than the average. That differs from the traditional picture of the AI wave as primarily hitting office work, and suggests that robotization will mostly affect the same groups that earlier waves of automation have already hit.
It is important to hold on to what the 74 percent figure does not mean: it does not mean robots will replace 74 percent of workers, or that 74 percent of occupations are threatened. The index measures how well robots can perform job tasks today — a degree of technical exposure. Whether exposure actually translates into layoffs depends on costs and other factors Anthropic's index does not capture. Anthropic's own findings on the language-model side support this distinction, as we will see.
The Context: The March LLM Study
The robot index comes half a year after Anthropic's first exposure study, which concerned language models. According to secondary coverage from Mezha, 127e3035, citing Anthropic researchers, that study found that computer programmers had the highest observed exposure — 75 percent of tasks covered — while data entry clerks stood at 67 percent, with customer support specialists also among the most exposed occupations.
Perhaps the most relevant finding from the LLM side for interpreting the robot index is this: the most exposed occupations saw no statistically significant increase in unemployment, according to the same coverage. Exposure, in other words, but no measurable layoffs — a pattern suggesting that technical coverage and actual labor-market impact are two different things for language models as well. The Mezha article is secondary coverage of a different, earlier study and provides context, not independent confirmation of the robot findings.
Open Questions
Several caveats come with the numbers. First, the entire study is self-published by Anthropic, the company behind the language model that scores the data. That does not make the findings wrong, but it means none of the key figures have been independently verified.
Second, parts of the methodology are unclear in the sources, particularly the cost estimates behind the 0.3 percent figure and the 40-year projection. Without full insight into how these are calculated, it is difficult to assess how robust they are.
Third, the fundamental question remains of whether the cost–capability gap actually holds over time. Anthropic's projection rests on historical price trends continuing. If robot prices fall faster, the gap could close far sooner than the 40-year horizon suggests. If they do not, robotization will remain a technical possibility more than an economic reality, occupation by occupation.
That last point may be the study's most important contribution: the debate about AI and work has largely been about the can question. Anthropic's robot index argues that the real question — at least for the coming decades — is at what price. And on that point, the answer, by the company's own figures, is: for very few.