Anthropic Says Robots Can Do 34% of Work, Cheaper for 0.3%

Anthropic’s study “What work can robots do?”, published on 30 September 2026, finds that robots today can perform 74% of physical tasks in the US, making up 34% of working hours. It also finds robots are cost-competitive with people for 0.3% of work.

The study resurfaced this week. A Polymarket post on X late on 7 October (UTC) said it found “AI & robots together are already capable of performing tasks covering 81% of U.S. employment.” The post now carries a Community Note pointing to the 0.3% cost figure.

What the Study Measures

The authors, Russell Legate-Yang and Maxim Massenkoff, start from O*NET, a database of around 900 occupations linked to around 19,000 job tasks. Claude scores each task as physical or not; 7,594 tasks are classified as physical.

For each physical task, Claude searches for specific robots and rates where one can do the work. The post says cited sources must show robot deployments, commercial sales or demonstrations, and that “only demonstrated robot capabilities count.”

The ratings use four tiers. E0: a robot cannot perform the task. E1: it can, in a purpose-built robotic environment such as a factory assembly line. E2: in a structured human workplace such as a logistics warehouse. E3: in an unstructured environment such as a city road.

Tasks are weighted by the number of workers who do them and by the share of working time spent on them, which the authors estimate with Claude.

Where the 81% Comes From

By working time, the post puts 54% of all US job tasks in cognitive and interpersonal work and 46% in physical work. Of all tasks, 12% cannot be done by any robot today, 23% can be done in purpose-built environments, 10% in structured workplaces and 1% in unstructured settings.

Adding the three exposed tiers gives the headline robot figure: 74% of physical work, or 34% of all work, “can be done by robots in some circumstances.”

The 81% combines that with an earlier measure of LLM exposure, the share of tasks for which an LLM could halve the time required. The post says “around half of work is exposed to LLMs alone, but that rises to 81% when considering robots.” Its key findings round this to about 80% of job tasks by working time.

Any robot tier from E1 up counts in the combined measure, so a task that a robot can do only on a purpose-built factory line is counted as exposed. The introduction phrases the same result as “Robots and LLMs together expose all but one-fifth of employment.”

The Cost Test

The study then asks whether robots are cheaper. For each exposed task, Claude estimates the annual cost of the robots cited, and the authors compare it with the worker’s total compensation for the time spent on that task.

On that test, robots are cost-competitive for 0.3% of job tasks. The post says that if robot price declines follow past trends, of roughly 3% a year since the 1990s, “it will take 40 years for that share to reach 10%.”

Packers and packagers are the largest occupation where robots already cost less: robots cost around $45,000 a year to replace one worker, against about $49,000 for the worker. Packers spend 97% of their time on tasks robots can do, the post says. The US employs around 560,000 of them. Robotaxis are estimated to cost around $7,000 more than taxi drivers.

Elsewhere the gap is wide. The robots needed to do welders’ tasks would cost around five times more than the welders, and cleaning robots are several times more expensive than dishwashers, janitors and cleaners.

Costs are not the only barrier the study lists. Capabilities prevent adoption for around 70% of physical tasks, and half would not be automated at scale without better manipulation. Regulation rules out 14% of physical tasks and human preferences hold back a quarter.

Who Is Exposed

Nine of the 10 most exposed occupations are vehicle operators, led by taxi drivers with an index of 2.2 out of 3. Nursing and general repair jobs are not highly exposed, the post says, because present-day robots can do little of that work even in controlled settings.

Workers in the top fifth of the exposure index are 55 percentage points less likely to hold a bachelor’s degree or higher than unexposed workers, earn around $30 less per hour, and face an unemployment rate more than twice as high.

The authors test the measure against history. In the appendix, an occupation whose tasks could all be done by robots had wages 7.1% lower 20 years later than an unexposed occupation in the same industry, and is predicted to lose 34.2% of its jobs over about 20 years.

The Structural Read

The robot measure counts capability, not deployment. A task a robot can do only on a purpose-built line counts as exposed at E1, and E1 is the largest tier in the study: 23% of all work time, against 1% at E3, in unstructured settings.

Cost is where the study separates capability from adoption. Holding tasks and wages fixed, at about 3% annual price declines robots are not cost-competitive for half of physical work until 2085, the study says; in its fast scenario, by 2050.

The exposure also falls on different workers. The study says robot-exposed jobs pay less and are more physically demanding, a pattern it calls in many ways the opposite of what is typically found for LLMs.

Three Implications

LOGISTICS AND PACKAGING OPERATORS Packers and packagers are the largest occupation where the study already finds robots cheaper: around $45,000 a year against about $49,000 for the worker.

ANYONE CITING THE 81% The study defines it as the share of US work time exposed to LLMs or robots in some setting today; its key findings round it to about 80% and set the 0.3% cost figure beside it.

ROBOTICS BUILDERS The study names manipulation as the main missing capability: half of physical tasks would not be automated at scale unless robots handle objects better.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The framework puts it this way: “The compounding is not uniform across capability dimensions.” And: “Picking the right axis matters more than executing well against the wrong one.”

As we read it, the robots study measures two axes: what robots can do in some setting, 34% of US work, and where they already cost less than people, 0.3%. The viral figure carried only the first axis, for LLMs and robots combined (81%).

What Is Not Established

The 81% and 34% figures are measures of what robots and LLMs can do in some setting today, not of what is deployed or adopted. The post calls its cost-competitiveness measure “more suggestive” given adoption frictions and cost uncertainties.

Task ratings and robot costs are Claude’s estimates using web search, and time shares are also estimated with Claude, as the post describes. The study’s timelines are for robots becoming cost-competitive, such as 40 years to reach 10% of work at past price trends, not for jobs being replaced.

The Bottom Line

Anthropic’s study puts 34% of US work within reach of today’s robots in some setting and 81% within reach of robots or LLMs, but finds robots cheaper than people for 0.3% of work. The 81% is the figure that went viral this week; the study’s own key findings put the 0.3% cost figure alongside it.