Anthropic just published one of the more useful pieces of research on how robots will actually change physical work, and the finding that matters most is simple: robots already have the physical skill to do a huge share of blue collar jobs in America, but almost none of it makes financial sense yet.
The researchers used their own AI, Claude, to go through nearly 19,000 job tasks pulled from a US government database covering around 900 occupations. For each task, they asked whether a real robot exists today that can do it, and under what conditions. The headline number: robots can already perform three quarters of physical tasks in some setting, covering a third of all working hours in the US economy.
That sounds alarming until you see the price tag. Robots are cheaper than a human worker for only about 0.3 percent of job tasks today. At the pace robot prices have fallen since the 1990s, roughly 3 percent a year, it would take about 40 years to push that share up to just 10 percent.
Which jobs sit at the front of the line. Nine of the ten most exposed occupations involve driving: taxi drivers, truck drivers, delivery workers. Warehouse packers and recycling sorters rank high too, since robots that scan, grip, and move boxes are already common and getting cheaper. Packers and packagers are the clearest case: robot costs already beat human labor costs for this job, and employment in the role has fallen 22 percent since 2015, with government forecasts putting it among the fastest shrinking jobs in the country.
Nursing, general repair, cleaning, and skilled trades tell the opposite story. A robot can weld a flat seam but can't yet climb a ladder, feel its way around a hidden pipe, or finish drywall with the touch of a human hand. The report points to hand and finger dexterity as the single biggest capability gap holding robots back, limiting half of all physical tasks across the economy.
One category sits in between: taxis. Robotaxis are already estimated to cost only about 7,000 dollars a year more than a human driver, putting them close to matching human cost. The real obstacle there isn't the robot, it's the rulebook. Cities and states are still fighting over where self driving cars are allowed to operate, so regulation, not technology, will decide how fast this job changes.
There's a useful historical check built into the report. Looking back 50 years, jobs that were more exposed to the robots of that era later saw bigger wage and job losses than less exposed jobs. That track record gives the researchers some confidence their measure isn't just a thought experiment.
It also lines up with a pattern Anthropic found in its separate research on office jobs and AI chatbots, where the gap between what AI could theoretically do and what workers were actually using it for ran as high as 50 to 65 percentage points. For robots, that gap is even wider, and the reason is straightforward: software is cheap to copy, hardware is not.
For any business built around physical labor, logistics, retail stocking, food production, or driving fleets, the lesson isn't that robots are coming for everyone. It's that a small number of roles, packing, sorting, driving, are already close to the point where a robot is cheaper than a worker. Everything else has years, in many cases decades, of room before the economics make automation worth doing.