Anthropic Says Robots Can Already Do a Third of US Work Hours — at 300x the Wrong Price

Anthropic's economics team has published a "robot exposure index" that ranks roughly 900 US occupations by how much of their physical work today's robots can actually perform. The headline finding is deliberately two-sided: robots can do 74% of physical job tasks, amounting to 34% of all US working hours, but they are cheaper than a human worker on only 0.3% of tasks. If robot prices keep falling at historical rates, Anthropic estimates it would take about 40 years for the cost-competitive share to reach 10%.

That gap between what machines can do and what anyone will pay them to do is the most useful thing in the report, and it is a corrective to a year of humanoid-robot demo reels.

How the index is built

The team starts from O*NET, the US government occupational database covering around 900 occupations and roughly 19,000 task statements. Claude scores each task description against a rubric for physical, cognitive, and interpersonal requirements; 7,594 tasks were classified as physical and then rated for robot exposure.

The exposure scale is about environments, not raw capability, which is a smart design choice. E0 means no robot can do the task at all. E1 means a robot can do it in a purpose-built setting like an assembly line. E2 means a structured human workplace such as a logistics warehouse. E3 means an unstructured environment such as a public road. Anthropic defines a robot narrowly as an autonomous physical machine that senses and acts — an automated car wash counts, a surgeon-controlled teleoperated surgical system does not.

To rate a task, Claude generates concrete examples of how the work is actually performed and how often, searches the web for specific robots, and must cite sources showing deployments, commercial sales, or demonstrations. Robots also have to do the work about as well as a person, accounting for reliability, error rates, and speed. A task's rating is set by majority rule: the least structured environment where robots can handle at least half of the task's examples, weighted by time.

The worked example in the report is instructive. "Dig trenches" comes out unexposed overall, because while an autonomous excavator retrofit can cut linear trenches in open ground (E3), most of the real work involves careful hand-digging around buried pipes and cables, which robots cannot do (E0). That is the kind of granularity most automation forecasts skip.

Where the work sits

Across all US tasks weighted by working time and employment, Anthropic puts cognitive and interpersonal work at 54% and physical work at 46%. Within the full task base, about 12% is E0, 23% is E1, 10% is E2, and roughly 1% is E3. In other words, half of all physical work that robots can do, they can only do inside environments engineered for them. Just 2% of physical tasks can be performed in genuinely unstructured settings, and most of that is driving.

The occupation rankings reflect that. Nine of the ten most exposed occupations are vehicle operators. Taxi drivers top the list at 2.2 on the 0–3 index; shuttle drivers and chauffeurs score 2.0, with Waymo vehicles cited as covering some tasks. Anthropic is careful to note that autonomous vehicles have not yet displaced driving jobs at scale.

Lower down, the shape of exposure varies in ways that matter for anyone modeling adoption. Drywall tapers score 1.6 because their work splits cleanly: 41% of their time is on tasks a drywall-finishing robot can do in unstructured settings, while 26% — the messy tape-and-mud embedding step — is unexposed. Recycling and reclamation workers score 1.7 for the opposite reason: over three-quarters of their time is on E2 tasks a vision-and-suction sorting robot can handle, with only 8% unexposed. Same rough score, very different automation paths.

The cost wall is the real finding

Strip out the capability story and the economics are stark. Anthropic says robots are cost-competitive with human labor on 0.3% of tasks. Robot spending remains about 1% of total US equipment investment, though the report cites business surveys suggesting US adoption could nearly double within three years.

After cost, the report lists capability as the next barrier — the dexterity to untangle wires is the example given — followed by human preferences and regulation. The source text available here cuts off before the full barriers analysis, so the relative weighting of those factors is not spelled out.

The 40-year extrapolation to 10% cost-competitiveness deserves a flag in both directions. It assumes historical price declines continue, which ignores the possibility that mass-manufactured humanoids break the trend; it also assumes nothing goes the other way on labor costs, integration expense, or financing. Extrapolations like this are scenario anchors, not forecasts.

Who is exposed, and how it differs from LLMs

Using the 2020–2024 American Community Survey, Anthropic compares workers in the top quintile of robot exposure with the roughly 20% of workers with zero exposure. The exposed group is 20 percentage points less likely to be female, 16 points more likely to be Hispanic, 55 points less likely to hold a bachelor's degree, earns about $30 less per hour, and faces an unemployment rate more than twice as high. BLS occupational requirement data show these jobs more often involve carrying weight, extreme heat, or hazardous contaminants.

This is close to the mirror image of LLM exposure, and that is the report's second argument. Combining robot exposure with LLM exposure measures from prior work, Anthropic estimates roughly half of work is exposed to LLMs alone, rising to 81% when robots are added. Transportation and moving jumps from under 15% to about 90%. Office and admin support approaches 100%. Protective service goes from 31% to 61%. The least-exposed groups — personal care and service, installation and repair, healthcare support, community and social service — share hands-on, face-to-face, and often regulated work. No occupation group falls below 40% combined exposure.

What the measure does not establish

The backtest is the most interesting validation attempt: Anthropic rated robot exposure at several points since 1977 and found that more-exposed jobs saw larger subsequent declines in wages and employment, controlling for industry trends. That is correlational evidence that the measure tracks something real, but a rubric applied retrospectively by a 2026 model to historical task descriptions is not the same as a pre-registered forecast.

Three other limits are worth holding onto. First, the entire exposure layer is Claude's judgment over web-sourced evidence, including vendor demonstrations — Anthropic says results are similar if demonstration-based ratings are dropped, which helps but does not eliminate marketing-grade claims as an input. Second, task exposure is not job loss; most of these jobs bundle exposed and unexposed tasks, and partial automation historically reshapes roles more often than it deletes them. Third, the index measures single-robot task coverage, not the systems integration, maintenance, and workflow redesign that dominate real deployment budgets.

What to watch

• Whether robot unit economics bend away from the historical price trend Anthropic extrapolates from, particularly for general-purpose humanoids now being tested in car plants and warehouses.
• Whether the business-survey signal of near-doubling US robot adoption within three years shows up in actual equipment investment, which currently sits near 1%.
• E2 tasks — work in structured human facilities, 22% of physical task time — as the most plausible next front, ahead of unstructured E3 deployments.
• Whether Anthropic releases enough of the underlying ratings and cited sources for independent researchers to audit Claude's judgment calls against field data.
• Regulatory and preference barriers in healthcare and personal care, where the report finds low combined exposure but does not quantify how much of that is capability versus rules.