1. What does Anthropic's robot report conclude?
Anthropic's research team has published a study estimating that robots can already perform about 74% of the physical tasks done in the United States, equal to roughly 34% of all working hours, but that they are cost-competitive with human labor for only about 0.3% of tasks. The report, titled "What work can robots do?" and released on September 30, 2026, frames automation as a question of price rather than technical ability.
The distinction matters because public debate about robots and employment usually centers on what machines are able to do. Anthropic's analysis argues that capability is only the first filter. A task a robot can perform in principle remains in human hands as long as a person can do it more cheaply, and for nearly all physical work that is still the case.
2. How was the robot exposure index built?
The authors scored roughly 7,600 physical tasks drawn from O*NET, the U.S. government's database that breaks occupations into individual tasks. They used Claude to rate each task against documented robot capabilities, citing sources for the ratings. The results were combined into a robot exposure index that measures how much of an occupation's work current robotics could handle.
Tasks were placed on a four-level scale. At E0, robots cannot do the task. At E1, they can do it in purpose-built settings, such as a factory line designed around the machine. At E2, they can work in structured workplaces. At E3, they can operate in unstructured environments such as public roads. Across U.S. working time, about 12% of tasks fall at E0, 23% at E1, 10% at E2 and 1% at E3, which together account for physical work making up about 46% of working time.
The researchers also tested the approach against history. A backtest starting in 1977 found that occupations more exposed to robots subsequently saw larger declines in wages and employment. The report adds that robots have been gaining the ability to perform about 2% of previously impossible physical tasks each year.
3. Why measure work by task rather than by job?
Studies of automation increasingly look at tasks instead of whole occupations, because a single job usually combines many different activities. A warehouse worker might lift boxes, scan labels, talk to supervisors and clear jams. A machine may be able to handle some of those duties and none of the others, so asking whether a robot can replace a job gives a less useful answer than asking which parts of the job it can perform.
The task-based approach is the reason the report can say that robots could technically handle about three-quarters of physical tasks while still leaving most workers in their roles. Capability is counted task by task and then weighted by the share of working time each task occupies, which is how the 74% figure converts into roughly 34% of all working hours once the non-physical half of the workday is included.
It also explains why the authors separate technical exposure from economic viability. Exposure describes what is possible. Viability asks whether using the machine costs less than paying a person, once purchase, maintenance and integration are accounted for. The report treats the second question as the binding constraint and measures it separately from the first.
4. Why does cost matter more than capability?
The headline capability figure comes with several qualifications. Most of the 74% exists only in controlled conditions: purpose-built or structured environments where the task can be standardized. Manipulation skills, such as dexterity and handling objects, are the most common gap that keeps robots from doing more. Regulation and human preferences also limit adoption in some areas, particularly healthcare, protective services and education.
Cost is the largest barrier. According to the report, robots are cheaper than people for about 0.3% of tasks. That slice covers roughly 300,000 workers whose duties are about 95% robot-doable and for whom a machine is already the lower-priced option. For everyone else, the economics still favor human labor, even where the technology could in theory do the work.
5. How long could it take for robots to become cost-competitive?
Anthropic modeled the pace of change using price declines. If robot costs fall by 3% a year, reaching cost competitiveness for 10% of tasks would take around 40 years and would require a cost reduction of roughly 70%. At that same rate, robots would become cost-competitive for half of physical work around 2085.
The report also considers a fast-adoption scenario in which prices fall more quickly. In that case, the half-of-physical-work threshold would arrive around 2050. The gap between the two dates shows how sensitive the outlook is to a single assumption, the speed at which hardware becomes cheaper, and why the authors treat the trajectory of prices as the variable to watch.
6. Why a cost gap can persist even when robots work
A robot that performs a task reliably is not automatically a better choice for an employer. Machines require capital spending, installation, upkeep and, in many cases, a workplace redesigned around them. Human workers, by contrast, can switch between tasks, adapt to unexpected situations and be hired in small increments. Those advantages are not captured by a capability score, but they weigh heavily in a purchasing decision.
This helps explain why the share of tasks where robots are already cheaper is so small even though the technical share is large. The 0.3% figure reflects work that is repetitive, well defined and performed in settings that suit machines, and where labor costs are high enough to justify the investment.
The report's timelines follow the same logic. They do not assume breakthroughs in what robots can do. They assume that prices decline at a steady annual rate, and they show how many decades of that decline are needed before the economics change for a meaningful share of work. That framing gives workers, employers and policymakers a measurable quantity to track instead of a vague sense that automation is coming.
7. Which workers are most exposed to robots?
Exposure is uneven. Nine of the ten most exposed occupations are vehicle operators, and taxi drivers rank first with an index of about 2.2 out of 3. Warehouse packers and packagers form the largest group for which cost-competitive robot alternatives already exist.
The workers in highly exposed jobs differ from those most affected by language-model automation. According to the study, they are more often male and Hispanic, less likely to hold a bachelor's degree and earn about $30 less per hour than workers in unexposed roles. Their unemployment rate is more than double that of unexposed workers. Exposed jobs also tend to involve carrying heavy loads and harsher working conditions.
At the other end of the scale, nursing and general repair work show low exposure, as do other interpersonal or hands-on fields such as healthcare support and personal care. Jobs that depend on social skills or on physical dexterity in unpredictable settings remain the hardest for robots to enter.
8. How do robots and AI language models compare in exposure?
The study also looks at both technologies together. Roughly 80% of job tasks, measured by working time, are exposed to either large language models or robots. The two overlap less than might be expected, because robots extend exposure into transportation and moving work, areas where language-model exposure is low.
That complementarity suggests that the groups most affected by software-based automation and those most affected by robotics are not the same, and that the two waves could reach different parts of the workforce at different times.
The gap of about $30 per hour between highly exposed and unexposed workers is large, and it indicates that robot exposure is concentrated in lower-paid work. The unemployment rate among exposed workers, more than double that of unexposed workers, shows that these groups already face weaker labor-market positions before any further automation arrives.
The pattern differs from the one seen with language-model exposure, which reaches into office and knowledge work. Robot exposure touches roles defined by physical effort, carrying and vehicle operation, so the groups most at risk are different. For policymakers considering retraining or support programs, the distinction suggests that tools designed for one kind of disruption may not fit the other.
9. What does the report mean for the automation debate?
The practical takeaway is that technical feasibility is a poor guide to near-term displacement. A machine that costs more than the worker it would replace remains a demonstration rather than a substitute. The 300,000 workers already in cost-competitive roles are a small share of the economy, while the 74% capability figure indicates how much physical work is technically within reach if prices fall.
Physical work is under half of working time, and cognitive and interpersonal tasks, where robots are weakest, make up the rest. Because exposure concentrates among lower-wage workers, any narrowing of the cost gap would fall most heavily on them, with transportation the first sector where it would be felt.
10. What to watch next
The key indicators are the pace of declines in robot prices and progress on manipulation, the capability the report identifies as the main technical bottleneck. A faster fall in costs than the assumed 3% annual rate would pull forward the dates in the report, while slower progress would extend them. Follow-up research from Anthropic or other groups that tests these estimates against real deployment data would show how closely the index tracks what firms actually adopt.
Readers should also watch how employers describe their own investment plans, since adoption decisions are made firm by firm and sector by sector. The report's central message is that the number to follow is not what robots can do, but the point at which doing it becomes cheaper than paying a person.