Anthropic Study: Robots Can Do 74% of US Physical Tasks but Beat Human Cost on Just 0.3%

TL;DR

Anthropic published What work can robots do? on September 30, a robot exposure index for the US economy built by economists Russell Legate-Yang and Maxim Massenkoff. The headline split: robots that exist today can perform 74% of physical tasks (34% of all US working hours) in at least some setting, but they are cheaper than the human doing the job for just 0.3% of work. If robot prices keep falling about 3% a year, as they have since the 1990s, it takes roughly 40 years for that share to reach 10%. Combined with LLMs, about 80% of job tasks by working time are now exposed to some form of automation.

What Anthropic measured

The study starts from O*NET, the US task database covering around 900 occupations and 19,000 task descriptions. Claude first sorted tasks into physical versus cognitive and interpersonal work, which flagged 7,594 tasks as physical: things like "Dig trenches" that no amount of software can do without a machine attached.

Claude then searched the web for real robots that can do each task and rated it on a four-step scale based on how controlled the environment has to be:

• E0: no robot can do it today.
• E1: a robot can do it in a purpose-built space, like an assembly line.
• E2: a robot can do it in a structured human workplace, like a logistics warehouse.
• E3: a robot can do it in an unstructured environment, like a city road.

Only demonstrated capabilities count, and every rating had to cite a deployment, a commercial sale, or a demonstration. The authors say results are similar if demonstrations are dropped. Tasks are weighted by how many workers do them and how much of their time each task eats, with employment from BLS 2025 data and time shares estimated by Claude.

The environment scale is the clever bit. A robot that only works on a factory floor and a robot that works on a public street are not the same threat to a job, and this index says so. Think of it as the difference between a chef who can only cook in their own kitchen and one who can cook in yours.

Who is exposed

Driving dominates. Nine of the ten most exposed occupations (with at least 20,000 jobs) are vehicle operators, thanks to autonomous cars, trucks, tractors, and pavers. Taxi drivers top the index at 2.2 out of 3, and shuttle drivers and chauffeurs score 2.0, with Waymo cited for their tasks. Warehouse work follows: robots like Amazon's touch-sensing Vulcan pick and stow merchandise, and recycling sorters have over three-quarters of their task time rated E2.

The workforce profile is close to the inverse of LLM exposure. Using 2020-2024 American Community Survey data, workers in the top fifth of the index versus unexposed workers are:

• 20 percentage points less likely to be female.
• 16 percentage points more likely to be Hispanic.
• 55 percentage points less likely to hold a bachelor's degree.
• Paid around $30 less per hour, with an unemployment rate more than twice as high.

Nurses and general repair workers sit near the bottom: present-day robots can do little of their work even in controlled settings.

Robots plus LLMs cover four-fifths of work

The study bolts the new index onto LLM exposure from the GPTs are GPTs research and Anthropic's own labor market measure from March. LLMs alone expose around half of work. Adding any robot that can do a task (E1 or higher) raises it to 81%. Office and admin support goes to nearly 100%, because the light physical tasks the chatbot could not do turn out to be robot-shaped.

What is left over is hands-on and face-to-face work. Personal care and service jobs sit around 40% exposed even with both technologies counted, alongside installation and repair, healthcare support, and community and social service.

The cost wall

This is where the report turns from alarming to boring, in the best way. For every exposed task, Claude estimated what it would cost a robot to produce a year of the same output, including annualized hardware and variable costs, and compared it with the worker's total compensation. A robot is only counted as competitive when it is cheaper.

• Packers and packagers are the largest occupation exposed to cost-competitive robots. The robot stack for their tasks costs over $2 million to buy and install but replaces the yearly work of around 14 workers. Employment in the job has already fallen 22% since 2015.
• Taxi drivers: robotaxis are estimated to cost only around $7,000 more than a driver, but face regulatory hurdles.
• Welders: AI welding robots exist, but the full set needed to do the job costs around five times more than the human.
• Dishwashers and janitors: paid $25,000 to $30,000 less than welders, yet their robot analogs are still several times more expensive.

In aggregate, robots are cost-competitive for 0.3% of all work, which still includes roughly 300,000 workers for whom robots can do 95% of their tasks. A 20% across-the-board price cut, about seven years at the historical 3% annual decline, would cover the physical work of 2.8 million workers, or 0.8% of all working time. Reaching 10% of work needs about a 70% cost decline, which is the 40-year figure.

What else blocks robots

Claude was also asked which barriers would stop robots from doing a big share of each physical task. Capabilities block around 70% of physical tasks, and manipulation is the standout gap: half of physical tasks would not be automated at scale until robots get better at handling objects. Planning and reasoning, the part AI is most likely to fix soon, limits only 8%. Regulation rules out 14% of physical tasks, mostly in healthcare, protective service, and education. Human preference holds back about a quarter, from "Dress children and change diapers" to greeting restaurant guests. Nearly every task also needs costs to fall.

Does the index predict anything?

The authors backtested it. They rated exposure for 1977 job tasks and found that jobs more exposed to the robots of the time saw wage and employment declines in later decades, even after controlling for industry trends. In 1977, robots could not do 62% of physical tasks; today's robots can do all but 24% of that same set, which works out to robots picking up about 2% of the remaining physical work each year.

Why this matters if you build things

• The method is reusable. An LLM with web search, a tight rubric, mandatory citations, and a released reasoning trail rated thousands of tasks. That is the same pattern you would use to size a market or audit a catalog, and Anthropic says its data release includes Claude's reasoning and cited sources for every rated task.
• Robotics pitch decks meet a cost table. "Robots can do X" is now cheap to claim and well documented. "Robots can do X for less than a person" is the claim that matters, and it currently holds for packing and close to holding for driving.
• Software still has the cheaper unit economics. LLM exposure arrives at software marginal cost. Robot exposure arrives as $2 million of hardware per packing line, which is why one shows up in your workflow this quarter and the other shows up in BLS tables over decades.

Caveats

The authors are direct that the rubric requires many judgment calls, that O*NET task statements are terse, and that the cost estimates are approximate. Task time shares and robot costs are themselves Claude estimates. The scenarios apply the same cost decline and capability gain to every task, while real robot makers chase specific markets first. And AI-powered robots could leapfrog the scale entirely, for example by learning to climb ladders or handle messy materials. Cost parity also does not mean mass unemployment: the authors cite work by Jones and Tonetti arguing automation boosts productivity only once machines are much cheaper than people.

Key Takeaways

• Anthropic's robot exposure index finds today's robots can perform 74% of US physical tasks, or 34% of all working hours, but mostly in purpose-built or structured settings.
• Robots are cost-competitive for just 0.3% of work; at 3% annual price declines it takes about 40 years to reach 10%.
• LLMs alone expose about half of work; LLMs plus robots expose 81%, with transportation jumping from under 15% to about 90%.
• Robot-exposed workers are less educated, earn about $30 less per hour, and face more than twice the unemployment rate of unexposed workers.
• Manipulation is the main capability gap, blocking half of physical tasks; regulation blocks 14% and human preference about a quarter.
• A 1977 backtest shows robot-exposed jobs saw wage and employment declines in later decades, which supports using today's capabilities as an early-warning signal.

Sources: Anthropic, What work can robots do?, full paper (PDF), appendix (PDF), Yahoo Finance, Digital Today, Anthropic, Labor market impacts of AI