On September 30, 2026, Anthropic's economic research team published a research post, “What work can robots do?”, along with a 56-page appendix. The team tackled two questions in a single dataset. How much of the work people do in the US can robots do today? And how much of that work can robots do for less than a person's labor cost? Put both numbers on the same base, total US work hours, and robots can do 34% of the work but are cheaper than people for only 0.3%. The work robots can do is more than 100 times larger than the work they can do cheaply.
Anthropic used Claude to check 19,000 tasks across 900 US occupations, one by one
The team started from O*NET, the US occupational information database. O*NET lists about 900 occupations, linked to about 19,000 task descriptions. First, Claude rated how physical each task description was and picked out 7,594 physical tasks. “Digging trenches” and “teaching dance” counted as physical. A typist checking totals on a calculator did not, even though the work uses the hands.
Next, Claude searched the web for real robots doing each task and quoted the source text verbatim. The team defined a robot as a machine that senses and moves on its own. So a car wash that scans the car to position its sprayers counted, while a remote surgical system fully controlled by a surgeon did not. Only real deployments, commercial sales, and demonstrations counted as evidence. According to the appendix, Claude Opus 5 ran about 650,000 web searches and left more than 275,000 quotes from over 90,000 web pages.
Take trench digging. Claude split the time spent on this task into several situations. 25% is digging a straight trench in open ground, and 20% is carefully digging by hand around buried pipes and cables. A robot can handle the first if an excavator is fitted with an autonomous digging control system. The second is beyond today's robots. The team counted a task as robot-doable only if the situations robots can handle made up at least half of the task's time, so trench digging was classified as a task robots cannot do. The public write-up does not say how the rest of the time was split.
Finally, Claude estimated what it costs for robots to do the work one person does in a year. It included the robot itself, installation, maintenance, software, electricity, supervision, and insurance, with fixed costs spread across years. This was compared with human labor cost: total compensation by occupation from the US Bureau of Labor Statistics (BLS), multiplied by the share of time the task takes.
In this study, Claude estimated everything: task classification, time share per task, robot capability ratings, robot costs, and the rollup to the occupation level. Human labor cost is also BLS total compensation multiplied by Claude's estimated time share. The post calls its cost estimates “approximate” and its results “more suggestive”. The appendix checked Claude's task-time and skill variables against BLS surveys and found correlations of 0.76 to 0.89. Neither the post nor the appendix reports a check of the robot capability ratings or robot costs against human expert judgment. Anthropic does not sell robots, but the post does not separately address the interest that comes with using its own model as the measuring tool.
Most of the work robots “can do” only works in spaces built for robots, like factories
Of the 7,594 physical tasks, robots can do 74%. Converted to the time people spend on them, these tasks come to 34% of total US work hours. But the post's summary adds a caveat: “mostly in limited settings.” The team sorted the environments where robots can do the work into four levels.
| Level | Where robots can do the work | Share of all physical tasks |
|---|---|---|
| E0 | Nowhere | About a quarter |
| E1 | Robot-only spaces such as factory assembly lines | About half |
| E2 | Structured human workplaces such as warehouses | 22% |
| E3 | Unstructured open spaces such as city streets | 2% |
A hospital delivery robot carrying medicine to a nursing station falls under E2. A robot on a conveyor assembling meal trays to match patients' diets falls under E1. Since about half of all physical tasks are E1, roughly two-thirds of the 74% that robots can do only works in spaces built for robots, like factories. Work robots can also do in open spaces like roads is just 2% of all physical tasks.
The range of what robots can do is widening slowly. According to the post, each year robots became able to do about 2% of the physical work they previously could not.
The Median, DataCamp's newsletter, argued that the 74% may be inflated. When Claude could not find a robot doing exactly the task described, it accepted robots doing similar tasks as evidence. The appendix describes this approach as well.
Packers are the biggest job where robots beat people on cost, and even there a robot saves only about $2,500 a year
Put what robots can do and what they can do cheaply on the same base, total US work hours, and the gap becomes clear. The post puts it this way: “Though robots can theoretically do tasks that add up to 34% of all work time, they are cost-competitive for just 0.3% today.”
Among jobs where robots are cheaper than people, the largest by headcount is packers. About 560,000 people do this work in the US, and they spend 97% of their time on tasks robots can do. The team spread the robot's fixed costs over a roughly 10-year life at an 8% cost of capital, then added maintenance, partial human supervision, and electricity, for an estimated robot cost of about $45,000 a year. Of the human labor cost of about $49,000 a year, the 97% share for robot-doable tasks comes to about $47,500. Compared with the robot's roughly $45,000, the robot is about $2,500 a year cheaper. Even in the biggest job where robots win on cost, the gap between robot and human is this small.
Welders sit at the other end, where robots cost far more than people. A robot with cameras and AI can do the welding itself on its own. But a human welder also moves large metal parts, climbs ladders to weld hard-to-reach seams, inspects quality, and handles grinding and finishing. Hand all of that to robots, and the robots needed cost about five times as much as a human welder. Janitors and dishwashers earn $25,000 to $30,000 less than welders, yet the robots for their work cost several times more than people. Cleaning is hard to standardize, and robots are slower than people.
The Median argued that the 74% may be inflated, and the same piece suggested the 0.3% could be off in the opposite direction. The 0.3% may understate how close robots have come to human costs. By the post's estimate, a robotaxi costs about $7,000 a year more than a taxi driver, so it falls outside the 0.3%. But next to welding or cleaning robots that cost several times more than people, the gap between a robotaxi and a driver is small. The post notes that robotaxis also face regulatory barriers beyond cost. Weighing the evidence on both sides, The Median concluded that the true figure could be a bit lower or a bit higher than 0.3%. On Reddit, one reaction to the 0.3% read: “humans being cheaper than robots is a pretty depressing form of job security.”
Even if robot costs fall 3% a year, it takes about 40 years before robots are cheaper for 10% of work hours
The post calculates that robot costs would need to fall about 70% for robots to be cheaper than people across 10% of total US work hours. At 3% a year, that takes about 40 years.
There are nearer-term numbers too. If robot costs fall 20% (about 7 years at 3% a year), robots become cheaper than people for physical work done by 2.8 million workers. On average, these workers spend 42% of their time on robot-doable tasks, and that time adds up to 0.8% of total US work hours.
The 3% rate came from weighing four sources together: the Bank of Japan's producer price index for robots, data from Danish firms, a US auto-welding robot case, and average selling prices of industrial robots in North America. All four are industrial robot data. The 40-year figure also assumes that costs fall equally across all tasks and that the rest of the economy stays the same when robots take on a single task. Neither the post nor the appendix separately addresses whether humanoids or AI-driven robots might get cheaper at a different pace.
The team checked whether scoring occupations by their share of robot-doable tasks lines up with historical data. Comparing occupations within the same industry using data from 1977 to 2024, occupations whose tasks robots could fully do had wages 7.1% lower 20 years later than occupations with no robot-doable tasks (95% confidence interval: 5.4% to 8.9%). In the same comparison, employment fell 34.2% (95% confidence interval: 16.4% to 52.0%). When the score was put into a regression alongside the Webb (2020) index, which identifies robot-doable work from patents, the declines in wages and employment showed up more strongly in this study's score than in Webb's index. But robots in this period mostly worked in factory-like E1 settings. So the appendix notes that these results say little about AI robots moving into warehouses and onto roads.
For robots to take over physical work, the technology is not enough. They have to get cheaper than people
The team also looked separately at what keeps robots from taking over tasks. The earlier 74% asked whether a robot can do a task in any environment. Here the question is what stands in the way of handing a task to a robot instead of a person. The 74% counts a task as doable even if it only works in dedicated spaces like factories, while the shares below count what blocks replacing people, so the two use different bases. Capability gaps block about 70% of physical tasks, and manipulation, the ability to grasp and handle objects, alone blocks half. Gaps in planning and reasoning, the abilities AI progress should improve most easily, block only 8%. Regulation blocks 14%, and people's preferences block a quarter. Dressing a child and changing diapers, or greeting guests and walking them through a wine list, are examples blocked by preference. Several factors can block a single task at once, so the shares for capability, regulation, and preference add up to more than 100%. After weighing all of these factors, the post adds one sentence.
“For nearly all tasks, costs would also need to fall.”
Put packers and welders side by side and the meaning becomes clear. Robots can do a packer's work, and they cost about $2,500 a year less than a person. Robots can also weld on their own, but handing a welder's whole job to robots costs about five times as much as a person. On technology alone, both sit inside the 74% of physical tasks robots can do. Robots actually start taking over the work only after their cost drops below human labor cost.
Until now, how much robots would change work has mostly been judged by whether robots can do it. Set the same work next to human labor costs, and robots arrive not across everything they can do, but first where they have become cheaper than people.
Further reading
What work can robots do? The original post from Anthropic's economic research team, with the exposure level table, robot costs by occupation, and charts of cost-decline scenarios. anthropic.com
Appendix to: What work can robots do? The 56-page appendix. It details how Claude found robots through web search, the validation against BLS surveys, and the 1977 to 2024 backtest. cdn.sanity.io
Robots Are Cheaper than People for 0.3% of Work DataCamp's newsletter The Median takes a close look at the study's methods. It lays out the evidence for why the 0.3% could be too high or too low. dcthemedian.substack.com