Anthropic finds robots can do 74% of US physical tasks but are cost-competitive on just 0.3% of work

Anthropic published research on 30 September 2026 estimating that robots are already technically able to perform 74% of the physical tasks in the US economy, but are cheaper than a human worker for only 0.3% of job tasks. The study, written by economists Russell Legate-Yang and Maxim Massenkoff, introduces a robot exposure index and argues that the cost of hardware, more than the reach of AI models, now sets the pace of physical automation.

The physical tasks that robots can handle account for 34% of US working hours, according to the paper. Once robots and large language models are counted together, about 80% of job tasks by working time are exposed to one or the other. Robot prices have fallen by roughly 3% a year since the 1990s, the authors write, and if that trend continues it would take 40 years for the share of tasks where robots are cost-competitive to reach 10%.

The team built the index from O*NET, the US occupational database that links around 900 occupations to around 19,000 task descriptions. Claude was used to judge whether a robot can carry out each task and in what setting. Tasks were sorted into four tiers: E0 when no robot can do the task, E1 when a robot can do it in a purpose-built cell such as a factory assembly line, E2 in a structured human facility such as a logistics warehouse, and E3 in an unstructured environment such as a city road. Ratings were weighted by working time and employment, and Anthropic released the model's reasoning and citations for each rating alongside a PDF of the full paper.

Taxi drivers are the most exposed occupation, with an index of 2.2 on the three-point scale, followed by shuttle drivers and chauffeurs at 2.0, a ranking driven by autonomous driving. Packers and packagers come closest to cost parity. The paper counts about 560,000 of them in the US, spending 97% of their time on tasks robots can perform, at a cost of about $49,000 a year per worker against about $45,000 a year for the robots needed to replace one person. Employment in the occupation has already fallen 22% since 2015. Welding sits at the other end: robots can weld, but automating the whole job would need equipment costing around five times as much as a human welder, the authors estimate.

Capability gaps remain the largest barrier, preventing adoption for around 70% of physical tasks, the study says. Regulation would not allow robots to perform 14% of physical tasks, and human preferences hold robots back for about a quarter of tasks. A backtest covering the 50 years from 1977 found that jobs more exposed to existing robots went on to see wage and employment declines in later decades, and the authors estimate that each year robots have become able to do about 2% of the physical work they previously could not.

The exposed workforce differs sharply from the rest. Workers in highly exposed occupations are 20 percentage points less likely to be female, 16 points more likely to be Hispanic and 55 points less likely to hold a bachelor's degree or higher, according to the paper. They earn around $30 less per hour and face an unemployment rate more than twice as high. Robots and language models also reach different work: fewer than 15% of transportation and moving tasks are exposed to LLMs alone, but about 90% are exposed once robots are included.

The findings land in a year when investors have priced general-purpose robotics on the opposite assumption. Figure AI raised $1 billion at a $39 billion post-money valuation in its Series C, and Skild AI raised $1.4 billion at $14 billion in January 2026, according to the companies. Walden Robotics, a Toyota research spinout, launched in July with a $300 million seed round at $1.1 billion. Those valuations rest on robots doing paid work at scale within years, while Anthropic's base case points to decades for cost parity across a meaningful share of tasks.

The central distinction in the paper is between what a robot can do and what it is worth deploying. A task may be technically feasible in a factory cell (E1) yet uneconomic once integration, supervision, slower cycle times and downtime are priced in. The welding example shows how the cost of a full job differs from the cost of its core motion. For robot makers, the implication is that the decisive metric is the cost per productive hour of a deployed system, not a demonstration of a new skill.

For buyers and policymakers the study narrows the near-term frontier to a small set of occupations where tasks are repetitive, time is concentrated in robot-capable work and wages are close to robot running costs. Packing and material handling fit that profile, which matches where warehouse automation and mobile robots have sold most widely. Interpersonal work and tasks demanding fine manipulation show the lowest combined exposure, which leaves care work and skilled trades largely outside the reach of current machines.

The method has limits the authors acknowledge. O*NET task descriptions are short and can omit details that matter for a robot, and the cost estimates are approximate. Using a language model to rate robot capability also invites questions about consistency, though the published reasoning allows outside researchers to audit individual scores. The 3% annual price decline is a historical average; a sharper fall in humanoid and component prices, driven by Chinese volume production, would shorten the 40-year horizon considerably.

The next tests come from the market. Robot makers' pricing for humanoids and mobile manipulators in 2027 contracts, US employment data for packers, material movers and drivers, and any independent replication using the released ratings will show whether the 0.3% figure moves faster than history suggests.