Robots Can Do 74% of Physical Tasks. Only 0.3% Pay Off.

The robot that can do your job and the robot that will do your job are two different machines, and Anthropic just measured the distance between them. In a study published September 30, economists Russell Legate-Yang and Maxim Massenkoff report that robots can technically perform 74% of physical job tasks in the United States. Then they ask what it would cost. On that test, robots beat human labor for just 0.3% of job tasks.

That second number is the one that will travel, and it should. Most forecasts about automation stop at capability: a robot has been shown welding, so welders are exposed. Anthropic's paper keeps going to the invoice, and the invoice changes the picture almost completely.

How the index was built

The researchers started with O*NET, the federal catalog of roughly 900 occupations and about 19,000 job tasks, and weighted each task by time spent and by how many people hold the job, using Bureau of Labor Statistics employment and wage data and the American Community Survey. Claude scored every task description against a rubric of physical, cognitive and interpersonal demands, and web searches were used to check what robots have actually been demonstrated doing and what they cost.

Physical tasks were then sorted into tiers by how much the environment has to be engineered around the machine. The lowest tier is a task no robot can do. Next is a purpose-built robotic cell, like factory assembly. Then a structured human facility, like a warehouse. The top tier is an unstructured setting, like a city street. According to the paper, only 2% of tasks are currently exposed in that last, messiest category, and 12% of all tasks cannot be automated at all.

Physical work accounts for 34% of all U.S. working hours, per the study. Within it, three-quarters of tasks are within reach of some robot somewhere, but mostly when the world has been rearranged to suit it. Anthropic's phrase for the requirement is "engineered environments."

The cost test

To get from capability to economics, the team had Claude estimate the annual cost of deploying a robot on a task, including hardware, installation, maintenance, energy and human oversight, and compared it with the share of a worker's total compensation spent on that task. A task counts as cost-competitive only when the robot is cheaper.

The clearest case where the math already works is packing. For packers and packagers, the paper estimates robot costs of about $45,000 a year against roughly $49,000 in labor, a margin of around $2,500. The occupation employs about 560,000 people, and its employment has fallen 22% since 2015. Welders and dishwashers sit on the other side of the line: robots can do the tasks, but the study puts them at three to five times the cost of a person.

The forward projection is sobering for anyone expecting a near-term wave. If robot prices keep falling at their historical rate of about 3% a year, the paper estimates it would take 40 years for cost-competitive tasks to reach 10% of work. Getting to that 10% requires roughly a 70% decline in costs.

What is actually holding robots back

Price is not the only wall. The study attributes half of blocked physical tasks to manipulation, the fine dexterity that hands do without thinking and grippers still fumble. Regulation keeps another 14% of physical tasks off-limits, concentrated in healthcare, education and protective services. Human preference blocks 25%, covering jobs like childcare and hospitality where people want a person in the room.

That mix explains the occupational ranking. The most exposed jobs are drivers. Taxi drivers score 2.2 out of 3 on the index and shuttle drivers and chauffeurs score 2.0, with other vehicle operators filling out the top ten, because autonomous vehicles are the rare robots that already work in unstructured environments. Nurses and general repair workers show minimal exposure. Personal care and service roles land around 40%.

Robots and language models hit different workers

The most consequential finding may be the comparison with Anthropic's earlier work on language models. LLM exposure runs through cognitive, office-based and higher-paying work. Robot exposure runs through physical, lower-wage jobs, with a median pay difference of about $30 an hour between the two groups, according to the paper. Robot-exposed workers are 20 percentage points less likely to be female, 55 points lower on the paper's education measure, and face twice the unemployment rate.

Put the two technologies together and the coverage is close to total. LLMs alone expose 50% of employment; adding robots lifts that to 81%, leaving roughly 19% of employment untouched by either. Transportation jumps from 15% exposure with LLMs alone to 90% once robots are counted.

Read the caveats before the headline

The authors are candid about the limits. Ratings rely only on demonstrated capabilities and cited deployments. O*NET task descriptions are terse and can hide complexity. Cost estimates are approximate, assume uniform price declines across very different tasks, and do not model wage responses or new jobs created. The paper also notes that AI-driven robots could "leapfrog" the exposure tiers if foundation models crack manipulation faster than hardware prices fall, which is precisely the bet that billions in humanoid funding are making this year.

There is an obvious tension in a lab that sells AI publishing a study that uses its own model to score which jobs AI and robots will reach. The method is transparent about that, and the conclusion cuts against hype rather than for it: capability is broad, economics is narrow, and the timeline is long.

The paper does not offer policy prescriptions. Its practical advice is to watch the occupations where the numbers already work, like drivers and warehouse packers, for early signs of disruption. For everyone else, the useful reframing is that the question is no longer whether a robot can do the task. It is whether anyone can afford to let it.