An Anthropic study suggests robots are decades away from replacing manual workers

A new Anthropic study titled "What work can robots do?" reaches a conclusion that runs headlong into the prevailing enthusiasm about humanoid robotics: robots are already technically capable of performing close to three quarters of the physical tasks that make up jobs in the United States, equivalent to 34% of all hours worked. That technical capability, however, runs into a far tougher economic barrier: today robots are cheaper than human labor for only 0.3% of existing work. The vast majority of those tasks also require highly structured, controlled environments, a far cry from the variability of the real world in a factory, a warehouse or a home.

The study builds its central argument on the historical trajectory of robot prices, which have been falling by around 3% a year since the 1990s. If that pace holds, the authors calculate it would take some 40 years for robots to become cost-competitive for just 10% of the work done in the United States. It is a projection that openly contrasts with expectations that have soared recently: at CES in January 2026, Nvidia chief executive Jensen Huang went so far as to predict an imminent "ChatGPT moment" for robotics, while Morgan Stanley had projected in May 2025 a $5 trillion humanoid robot market by 2050.

To test the optimism of robotics advocates, Anthropic's study also models a "fast" scenario in which costs fall up to four times more quickly than the historical trend and robots learn new tasks at twice the usual speed. Even under that favorable assumption, robots would not become cheaper than people for half of today's physical tasks until 2050. As the report itself notes, "robots would need to sustain record-rate price declines and quality improvements over the coming decades to enable rapid physical automation".

Beyond cost, the study identifies technical capability as the second major obstacle: tasks demanding fine dexterity, such as untangling cables, remain out of practical reach for today's robots. Human preferences and regulations further limit adoption across a significant share of tasks. Apollo chief economist Torsten Sløk backed these conclusions in an emailed note, warning that "even under aggressive projections, widespread displacement of physical labor will take decades", and adding that "severe bottlenecks in fine manipulation capabilities, regulatory constraints and human preferences will continue to significantly slow automation in the near term".

The picture changes, and grows more uncertain, when it shifts to office jobs, which are far more exposed to language models (LLMs) than to physical robotics. Anthropic chief executive Dario Amodei himself has repeatedly warned that artificial intelligence could eliminate up to half of entry-level white-collar jobs before 2030. Yet the company's own researchers are more cautious: a March paper by Anthropic economists Maxim Massenkoff and Peter McCrory found "no systematic rise in unemployment among the most exposed workers since late 2022", although it did find "suggestive signs that hiring of younger workers has slowed in exposed occupations". In an essay published on X in July 2026, McCrory concluded: "I don't expect unemployment to be notably higher a year from now, at least not because of AI".

The new study, also co-authored by Massenkoff, estimates that roughly half of work tasks are exposed to language models such as Claude, in the sense that a tool of this kind could halve the time needed to complete them. The authors caution that this metric measures what tools can speed up, not necessarily which jobs will end up being replaced. And they stress a key asymmetry between software and hardware: cost is not the same barrier for LLMs as it is for physical robots, suggesting that disruption in office work could arrive faster than in manual work. "Robots built from expensive hardware are not copied or distributed as easily as software or language models", the authors note, insisting that it is worth distinguishing between exposure to AI capabilities and exposure to robotic automation that is also cost-competitive.

The report also explores what kind of work currently lies beyond the reach of both LLMs and robots. It points to care and personal service jobs, where around 40% of tasks are exposed, but where face-to-face social interaction and physical contact with other people — areas in which both language models and robots struggle — remain central. Occupations such as installation and repair, healthcare support, or community and social services fit the same pattern, with relatively low exposure levels. Overall, the study sketches a future of automation far more gradual and uneven than the most enthusiastic robotics headlines suggest, with physical work protected for decades by cost economics, while office work faces a potentially faster transition, though still without clear evidence of mass job destruction.