Robots can already do three-quarters of the physical tasks in American jobs, but they are cheaper than a human worker for almost none of it. That is the central finding of a research paper Anthropic published on 30 September, which ranks occupations by how much of their work a machine could handle today.
Anthropic counts a robot as any autonomous physical machine that senses and acts, so a car wash that scans each vehicle qualifies and a surgeon-controlled instrument does not. Using that definition, the company says robots can perform 74% of physical tasks, which adds up to 34% of all working hours in the US. Cost is the catch. Robots are cheaper than people for just 0.3% of job tasks.
The paper rates each task by how messy the setting can be. A robot that works only on a purpose-built line, like a factory assembly station, scores low. One that copes with a warehouse scores higher. One that handles a city street scores highest. Today, robots manage only 2% of physical tasks in open, unpredictable surroundings, and most of those involve driving. Taxi drivers top the exposure ranking, and nine of the ten most exposed occupations are vehicle operators.
The cheapest case Anthropic found is packing. Its estimate for a set of robots that pack, move, inspect, box, label, and seal goods is more than $2 million to buy and install, replacing the yearly work of around 14 people. Spread over a ten-year service life, with an 8% cost of capital, the company puts the running bill near $45,000 per displaced worker. Packers and packagers cost employers about $49,000. The US employs roughly 560,000 of them, and Anthropic notes the occupation has shrunk 22% since 2015.
Most other jobs are nowhere near that threshold. Anthropic estimates that the robots needed to cover a welder’s full range of duties (positioning heavy parts, climbing to awkward joints, checking quality, grinding) would cost about five times what human welders do. Robotaxis come out only about $7,000 a year more expensive than drivers, but the paper says regulation stands in the way there.
The timeline is the sobering part. Industry and government figures cited by Anthropic show robot prices falling roughly 3% a year since the 1990s. At that pace, cost-competitiveness for 10% of human work needs a price drop of about 70%, which means around 40 years. The paper’s fast scenario assumes quality-adjusted prices drop at as much as four times the current pace while robot skills grow twice as quickly. Under it, robots become cheaper for half of physical work by 2050.
Anthropic also tested whether exposure predicts anything. It rated tasks as they stood in 1977 and compared them with later outcomes. Jobs more exposed to the robots of the day saw wages and employment fall in the decades after, even after adjusting for industry trends. Highly exposed workers differ sharply from the rest: the share with a college degree sits 55 percentage points lower, and hourly pay runs roughly $30 under the norm, according to the company’s analysis of census survey data. That is close to the reverse of the profile that studies of chatbot exposure tend to find.
Two limits deserve a flag. The ratings and cost estimates came largely from Claude itself, which scored tasks, searched the web for matching robots, and estimated prices, and Anthropic says the rubric requires many judgment calls. This is also the company’s own research, with no outside replication yet. The authors describe the cost figures as suggestive, not as forecasts, and note that humanoid robots or cheaper manufacturing could break the 3% trend.
Pairing robots with language models widens the picture considerably. Anthropic puts chatbot-style exposure at about half of all work, and the combined figure at 81%. Office and administrative jobs climb to nearly 100% once physical tasks are counted.
For anyone planning workforce or capital budgets, the practical number is $2,500: that was the annual saving in the one big occupation where Anthropic found robots already cheaper, and it is thin enough that one wage or borrowing shift could flip it.
Reported by Anthropic in its research paper “What work can robots do?”, published 30 September 2026.
