Anthropic says cheap robot labor is 40 years away. But ... it's pricing the wrong robots

Are useable humanoid robots really 40 years away?

Anthropic just told the humanoid robot industry it's about 40 years early.

Robots can already do 74% of the physical work Americans get paid for, according to a new Anthropic study published September 30 by economists Russell Legate-Yang and Maxim Massenkoff. But the big problem: today they're cheaper than the people doing it for just 0.3% of those jobs. And since robot prices have only fallen about 3% a year since the 1990s, the researchers say it'll take roughly 40 years before robots are cost-competitive for even 10% of human work.

It's careful, impressive research. 

I also think it's answering the wrong question.

The problem is which robots Anthropic is pricing. The researchers had Claude find specific robots that can do each of roughly 19,000 job tasks today, estimate what those robots cost, then total up the machines needed to cover each occupation. That's a reasonable way to price automation as it exists. But automation as it exists is a kind of a parade of specialists: a case packer here, a box erector there, a labeler, a sealer, an inspection station. 

So in Anthropic's model, automating a warehouse packer's job takes six kinds of machines and more than $2 million.

Shocker, that's not cost-competitive.

To be precise, Anthropic doesn't claim a robot can only ever do one task. Its occupation-level estimates explicitly try to avoid double-counting machines that cover several tasks. But three assumptions do most of the heavy lifting behind that 40-year number:

1. Robot capabilities are those of machines demonstrated today. 
2. A robot's cost is weighed against the labor of a single occupation. 
3. And the future follows the past: robots get about 3% cheaper a year, and learn to do roughly 2% of the tasks they previously couldn't each year (the paper's fast scenario doubles or quadruples those rates, and applies them uniformly to every task).
   

Every one of those assumptions describes the specialized-machine era. None of them describes what the humanoid industry is actually building: general-purpose AI running on general-purpose hardware.

If that works, a robot isn't a line item in one job. 

It's a worker that moves between jobs — unloading trucks at 6 a.m., tending machines at noon, packing orders in the afternoon, cleaning overnight — and picks up new skills through software updates. Its cost gets spread across every hour and every task it can do, across occupations, not just within one. And unlike a case sealer, its value goes up after you buy it.

That doesn't follow a 3%-a-year curve.

Anthropic sees some of this coming. The authors say humanoids or new manufacturing methods could drive costs down faster, and their fast scenario pulls the date robots become cost-competitive for half of physical work from 2085 to 2050. 

But even that is the old curve running faster, not a different curve.

My bet: we'll see genuinely interesting general-purpose robots doing real work in 2027, when Agility's Digit 5 ships and Figure, AgiBot and others push beyond pilots. (I know humanoids have had jobs for years, but I'm not 100% convinced they would have those if it weren't for piloting/novelty/innovation/test rather than just super-cold hard ROI calculations.) 

Robots that are commercially viable at scale, where a CFO signs a fleet order on ROI alone, will probably take a bit longer than that. But we're talking years, not four decades.

Anthropic's own data, read closely, makes the case.

What Anthropic actually measured

The study starts with O*NET, the U.S. Department of Labor's database of roughly 900 occupations and 19,000 job tasks. Anthropic had Claude, its own AI model, sort out which tasks are physical (7,594 of them), then search the web for specific robots that can do each one and score them on a four-level scale based on how controlled an environment the robot needs.

1. E0 means no robot can do the task. 
2. E1 means a robot can do it, but only in a purpose-built robotic environment like an assembly line. 
3. E2 means a robot can do it in a structured human workplace like a warehouse or hospital. 
4. E3 means a robot can do it out in the messy, unstructured world, like on a city street.
   

Look at the distribution and the 74% headline gets a lot less dramatic. About half of all physical work by time is E1, and another 22% is E2. Only about 2% reaches E3, and most of that is driving. In fact, nine of the 10 most robot-exposed occupations in the study are vehicle operators, with taxi drivers at the top of the list thanks to Waymo and friends.

That distribution tells you a lot about how automation has worked for 60 years. 

Industrial robotics mostly succeeds by making the world easier for the robot. You bolt it to the floor, fence it off, send identical parts past it on a conveyor in the same orientation every time, control the lighting, and keep people away. That can be astonishingly productive. 

But you're not really building a robot that works in a human world. You're building a little world in which a robot can work.

The humanoid pitch is the kind of the inverse of that: stop rebuilding the world for robots, and build robots for the world we already have. Human-height shelves, human tools, human doors, stairs, carts and workstations. Interestingly, Anthropic's scale bakes in that same logic. Robots "that work like humans count more toward exposure" than robots that need the workplace rebuilt around them, the authors write.

Which is exactly the bet humanoid makers are placing.

The $2 million packer (and the humanoid hiding inside)

The single most useful example in the report is packers and packagers, the largest occupation where automation/robots are already cost-competitive with people.

Automating that job takes six different kinds of machines, according to Anthropic: case packers, material-handling robots, machine-vision inspection stations, box erectors, print-and-apply labelers and case sealers. All in, that equipment costs more than $2 million to buy and install. Spread over a roughly 10-year life at an 8% cost of capital, plus maintenance, energy and part-time human supervision, it replaces the annual output of about 14 workers. That pencils out to around $45,000 per worker-equivalent per year, against about $49,000 in total compensation for a packer. 

Robots win by about $2,500 a year.

That's conventional automation at its best: high volume, highly structured, lots of capital, and enough throughput to amortize all of it. And it shows up in the labor data. There are about 560,000 packers and packagers in the U.S., employment in the occupation has fallen 22% since 2015, and the Bureau of Labor Statistics projects it will shed the 11th-most jobs of any occupation by 2035.

Now dig into the footnotes. The material-handling robots Anthropic cites for that $2 million line are Boston Dynamics' Stretch and Agility's Digit.

So there's already a humanoid in Anthropic's model of a cost-competitive automated job, doing one of the six jobs.

Here's the whole question for the humanoid industry in a nutshell: what happens when it does three of them? Or five?

For most other physical jobs, the gap is enormous. Robots that can do the work of welders cost about five times more than the welders, according to the report, because welding is surrounded by a bunch of other tasks: positioning heavy parts, climbing ladders to reach joints, checking quality, grinding and finishing. Dishwashers and janitors earn $25,000 to $30,000 less than welders, but their robot equivalents still cost several times more than they do, partly because cleaning robots can be up to 10X slower than people. To match what a U.S. Postal Service mail carrier does in a year, robots would cost about $166,000, against median compensation of $82,000. 

Robotaxis, by contrast, cost only about $7,000 more than taxi drivers. Their problem is regulation, not price.

There's also a lovely little example about archivists. Robots can already do one of their physical tasks, copying records to digital formats, for about $5,500 a year, versus about $6,800 worth of an archivist's time. Cost-competitive, right? Nope:  that task is about 8% of the job, so nobody is replacing archivists with a scanning robot.

That's the distinction most automation math struggles with. 

Spreadsheets treat jobs as collections of tasks. People experience jobs as a continuous stream of whatever needs doing next: move the linens, wipe up the spill, carry in the delivery, notice the broken fixture and tell maintenance, help the guest with their luggage. Humans are economically efficient partly because they're general-purpose machines that switch between tasks essentially for free, and they do a bunch of things nobody ever wrote down in a job description.

Anthropic knows this. 

The researchers note that simply adding up per-task robot costs could double-count robots or miss coordination costs, so they had Claude estimate total automation costs at the occupation level, net of redundancies. That's sensible. But it can only price the robots that exist today, and the robots that exist today -- especially, that you can buy today -- are overwhelmingly specialists.

Why not do the math on an actual humanoid?

So let's price a generalist.

Agility, which is going public in a $2.5 billion SPAC deal, gave a rare look at humanoid economics in its S-4 filing. In one illustrative model, a customer pays about $200,000 for a Digit 5, plus a $20,000 deployment fee and $36,000 a year for software and maintenance. Over five years, that's about $400,000, or roughly $80,000 a year. 

(Agility stresses the figures are examples and that real contracts are negotiated individually.)

Run that robot for one eight-hour shift and it's a terrible deal: $80,000 a year to replace a $49,000 packer, and that's assuming the robot is just as productive as the person, which it isn't.

But Digit 5 isn't designed to work one shift. It runs about 90 minutes on a charge, then charges in about nine minutes, a 10:1 run-to-charge ratio that Agility says enables more than 20 hours of work a day. Call it 20 hours a day, five days a week, 50 weeks a year, and you get 5,000 working hours annually. That's about $16 an hour for the robot, versus about $24.50 an hour for a packer earning $49,000 in total compensation (wages + benefits + costs) over 2,000 hours.

Which means that at roughly two and a half shifts, Digit 5 only has to be about 65% as productive as a human per hour to beat the cost of the person. At two shifts, it needs to be about 82% as productive. At one shift, it would need to be 163% as productive, which isn't happening anytime real soon.

At 70% of human productivity and 20-hour days, Digit 5 costs about $45,700 per worker-equivalent per year. That's almost exactly what Anthropic calculated for its $2 million multi-machine packing line. Same economics, very different shape: one machine you can walk to a different part of the building tomorrow, versus six machines bolted into a process.

That's my math, not Agility's or Anthropic's, and the caveats are significant. The pricing is illustrative. The 20-hour day is a spec, and Digit 5 doesn't ship until the first half of 2027. Two and a half shifts of utilization assumes there's actually two and a half shifts of work for it to do. Productivity per hour is the big unknown. And none of this counts the human who has to supervise the robot and rescue it when something goes wrong.

(Because, let's be honest, something will go wrong.)

Also, to be fair to Anthropic, its method already annualizes robot costs against human output, so a packing line that runs around the clock gets credit for that. Utilization alone doesn't break the model.

Breadth does.

Agility's pitch for Digit 5 is that it moves beyond the tote-shuffling that Digit 4 has done for nearly three years (65,000-plus operational hours across customers including GXO, Schaeffler, Amazon and Toyota) into depalletizing, machine tending, kitting, sequencing, inspection and palletizing. Each of those is a separate machine purchase in a conventional automation project. In a humanoid deployment, each one is a skill added to hardware you already own.

That's the property Anthropic's task-by-task approach can't really capture: a robot whose economics improve after you buy it.

But hey ... Digit 5 is the floor, not the ceiling

And Digit 5 probably isn't the most capable humanoid on the planet.

That's not a knock on Agility. Digit is arguably the most commercially proven humanoid in the world, with nearly three years of paid work behind it and a safety system built for exactly the environments where these robots need to operate. 

But ask most people who follow the industry closely which robots are the most advanced, and Digit usually isn't at the top of the list. Figure is probably further along on AI and dexterity: Figure 03 was redesigned around its Helix vision-language-action model, its fingertip sensors can register forces as light as three grams, and Figure says renting homes to train on chores delivered a 6X jump in success. (And Figure 04 is coming soon.) The top-end Chinese platforms move faster, and their makers iterate quicker than anyone. Boston Dynamics' production Atlas has 56 degrees of freedom, lifts up to 110 pounds, swaps its own batteries, and can learn a new task in under a day, according to the company. It also just got impressive new hands: four fingers (no pinky) with 13 degrees of freedom, a deliberate compromise between dexterity and durability, both of which matter for a robot expected to work 20-hour days.

Europe has serious contenders too. 

Agile Robots in Munich spun out of the German Aerospace Center's robotic hand research, has put more than 20,000 automation systems into factories, and builds a 71-degree-of-freedom humanoid, Agile ONE, with dexterous hands. When I toured its factory, two production lines building robot arms and torsos were entirely automated. Neura Robotics, an hour away by autobahn, raised up to $1.4 billion in June and reportedly lists its 4NE1 Gen 3.5 at about €98,000, with its own high-speed electric actuators.

Then there's speed, the single most important variable in my Digit math.

London-based Humanoid told me its wheeled HMND 01 is now performing some core tasks at roughly 80% of human speed and success rate, up from about 60% not long ago. It expects to hit and eventually beat 100%, but getting past human level will take reinforcement learning, not just copying people. 

"If we're only to use imitation learning, based on the data we collect with humans, then human performance is the limit," chief product officer Sotirios Stasinopoulos told me.

Plug that into the same math. At 80% of human productivity, a robot priced like Digit 5 and running 20-hour days costs about $20 per human-equivalent hour, comfortably under a packer's $24.50. A robot that costs 50% more than Digit 5 (about $600,000 over five years, or $120,000 a year) breaks even at full human speed. At 120% of human speed, it costs about $20 an hour and wins outright, even with the bigger price tag.

That's the point: capability is a cost lever. 

A faster robot amortizes the same hardware over more output. A more dexterous robot can take on more of the tasks in a job, which matters a lot when Anthropic says poor manipulation is what's blocking half of all physical work. A smarter robot needs less human babysitting, one of the biggest hidden costs in Anthropic's estimates. 

Better robots don't just make better demos. They pull break-even forward.

Even better, much of that improvement is software. Speed gains from reinforcement learning and new skills from a model update land on robots customers already own. A packing line bought in 2026 is still a 2026 packing line in 2031. A general-purpose robot bought in 2026 should be a much better robot by 2028.

Former Apple executive Rory Sexton, now at Agile Robots, put it plainly when I visited: "You need flexible robots that can do multiple things."

Caveats apply, of course. Few of these robots have Digit's commercial track record yet, and we don't know what most of them actually cost. Atlas's entire 2026 production is spoken for by Hyundai and Google DeepMind. Neura's flagship is only now starting to ship. Figure's deployments are still relatively small. 

Advanced isn't the same as deployable, and plenty of brilliant robots have never paid for themselves.

But if a less advanced robot already roughly pencils out on paper, the more advanced ones don't need to get much cheaper to beat it. They just need to ship, and keep working.

Three ways to bend a 40-year cost curve

For robots to be cost-competitive for 10% of today's work, Anthropic says robot costs need to fall about 70%. At 3% a year, that's 40 years. To do it in three years, costs would have to drop about 33% a year. In two years, about 45% a year.

That's a massive difference. But the humanoid industry has more than one lever to pull, and the levers multiply rather than add.

The first lever is price. Anthropic's own fast scenario has quality-adjusted costs falling up to four times faster than history, or about 12% a year. On price alone, that gets you to the 70% reduction in roughly a decade, not 40 years. But the early humanoid data is moving faster than that. IDC says humanoid makers shipped almost 25,000 robots in the first half of 2026 for revenue of more than $740 million. 

That's an average of about $29,600 per robot, down from about $37,000 a year earlier: a 20% drop in one year. Goldman Sachs says humanoid manufacturing costs fell 40% in a single year thanks to scale and commoditized components.

Neither of those is quality-adjusted, and the IDC average is skewed by a flood of small, cheap research and education robots. But if something like a 20% annual decline held, the 70% reduction would take about five and a half years. 

At 40% a year, about two and a half.

The low end of the market is getting absurd. 

Unitree's R1 starts at $4,900. A small San Francisco team is selling the wheeled Nori A3 for $1,688. PrimeBOT is selling home robots in China starting at about $2,750. Feather Robotics just launched a $29,990 bimanual robot with two seven-DoF arms, a wheeled base and 10 hours of battery life, and it already has more than $1 million in revenue. The company says robotics has been stuck between roughly $100,000 systems and "low-cost robots not meant to survive deployment."

Manufacturing scale is arriving too. 

UBTECH started production on September 12 at a Liuzhou factory designed for more than 10,000 humanoids a year, finishing one roughly every 10 minutes, with its own Cruzr robots doing palletizing, loading and material transport on the line. Agility's RoboFab in Salem, Oregon is designed for up to 10,000 units a year. Figure, which was doubling deliveries every month this spring, says Figure 03's components cost 90% less than its predecessor's.

The second lever is utilization. Humans work shifts. Capital equipment doesn't have to. A robot that works two and a half shifts cuts its cost per unit of output by 60% compared to one that works a single shift, with no change in sticker price. That's why fast charging and battery swapping matter so much: Digit 5's nine-minute charge, UBTECH's Walker S2 swapping its own batteries, Feather's hot-swappable packs.

Combine the two: 2.5X utilization plus a 25% drop in price gets you to a 70% lower cost per hour of work. That's Anthropic's entire 40-year gap, on paper.

"On paper" is doing a lot of work in that sentence, of course.

Reality is always tougher. In theory, as the saying goes, there is no difference between theory and practice.

The third lever is breadth. A specialized machine gains value only when it gets cheaper. A general-purpose machine also gains value when it learns something new. Suppose version one of a robot can economically do two hours of useful work a day, and a year later, with no hardware change, software updates get it to eight. 

Its effective cost per useful hour just fell 75%.

You can see early versions of this. Figure's first BMW deployment put Figure 02 robots in the body shop in Spartanburg for 11 months, where they helped build more than 30,000 BMW X3s. Figure 03 is now doing something different at the same plant: logistics sequencing. Hexagon and Schaeffler are taking a fleet approach, with Schaeffler planning to deploy at least 1,000 AEON humanoids across its global factories by 2032 and training them in a dedicated "Humanoid Gym" before they go into production. Every new skill that fleet learns raises the return on robots Schaeffler already paid for.

Anthropic makes a related point near the end of its paper: robot makers may prioritize automating work in their own factories first, much as AI companies prioritized coding agents because coding was useful to them. Robots building robots at UBTECH, Agility saying Digits will help build Digits, Tesla planning Optimus for its own lines: some of that is marketing, sure. But your own factory is also the cheapest possible place to collect training data and prove ROI.

The bottleneck is hands, not brains

There's another finding in the report that should get significant attention.

Anthropic asked which barriers would stop robots from doing each physical task at scale. Capabilities matter for about 70% of physical tasks. Regulation blocks about 14%, and human preferences (would you trust a robot to change a diaper?) hold back about a quarter.

But when the researchers broke down the capability gap, one skill dominated. Inadequate manipulation blocks automation for about half of all physical tasks. 

Planning and reasoning, the stuff AI is improving fastest, limits just 8%.

That's striking given where the money is going. Billions are flowing into vision-language-action models, robot foundation models and world models, and all of that matters. But Anthropic's analysis implies that a brilliant robot with mediocre hands hits a wall surprisingly fast. It knows the wire needs untangling. It just can't untangle it. It understands the bag needs opening, but can't get two compliant fingers onto the edges.

So the hand race may matter as much as the brain race. 

Figure 03's fingertip sensors can register forces as light as three grams. Agility is building swappable grippers for Digit 5 with ISO-standard mounts so customers can match end effectors to jobs. Impressive new hands are coming almost every week these days. And not all tasks require 5 fingers: AgiBot's wheeled G2 is working a live electronics production line at Longcheer with customized grippers.

Why 40 years might still be right

Now for the cold water, because there's plenty of it.

Most humanoids still aren't working. According to IDC, research and education, performances and exhibitions, and government data centers took 69% of humanoid shipments in the first half of 2026. That's down from 83.8% in 2025, which is real progress, but more than two-thirds of humanoids still aren't doing productive work in any traditional sense. Many of the ones in China are going to government-run robot schools that exist to generate training data. AgiBot leads the market with about 8,600 units in the half, and Unitree follows with roughly 5,000, but nobody's labor market is shifting on those numbers yet.

Reliability is brutal in manufacturing.

A robot that succeeds 95% of the time sounds great until it performs 1,000 actions in a shift and fails 50 times. Even 99% means 10 failures a day, each one needing a human. Useful autonomous work probably needs another nine, or two.

Supervision is a hidden cost. If one person can monitor and occasionally rescue 20 robots, the economics work. If every robot needs a dedicated teleoperator for big chunks of the day, you've built a very expensive remote-control system. Anthropic's cost estimates already include human supervision, exception handling and repair, which is part of why the numbers are so high.

Safety gates everything. Deploying robots without cages around people is what makes the multi-task humanoid model possible, and it's also what makes customers nervous. Safety, an Agility spokesperson told me via email this week, is "the long pole in the tent" for customers deploying fleets of humanoids, since it's what their health, safety and insurance teams need settled before they'll sign.

And cost parity isn't the finish line anyway. Anthropic notes, citing economists Daron Acemoglu and Pascual Restrepo among others, that automation tends to boost productivity only once machines are much cheaper than people, not merely as cheap. A humanoid that matches a packer's cost by a few thousand dollars a year is an interesting pilot, not an economic transformation.

The authors' bottom line is blunt: rapid physical automation would require robots to sustain record rates of price decline and quality improvement for decades. That's possible. But there's no law of physics that requires it.

So, 40 years or three?

Probably neither, taken literally. Forty years is a projection of historical industrial-robot economics. Three would require an almost unprecedented collapse in the effective cost of robot labor.

But the humanoid industry's core claim is that we're approaching a point where historical robot economics stop being a useful guide, because the thing being sold is changing. If what comes out of Agility, Figure, AgiBot, UBTECH, Hexagon, Unitree, Apptronik, Tesla and the rest turns out to be expensive single-task machines, Anthropic's timeline will look remarkably prescient.

If they turn out to be general-purpose machines that unload a truck in the morning, tend a machine at noon, inspect parts in the afternoon and palletize shipments overnight, then learn a new job next month via a software update, the question changes. We're no longer asking how fast industrial robot prices fall. We're asking how fast robot labor gets cheaper.

Those aren't the same curve.

AgiBot's Dr. Yao Maoqing told me in May that the industry's tipping point won't come from demos or production numbers, but from robots proving they deliver productivity in live environments over time. That's the right test. I think 2027 is the year we see general-purpose robots that are genuinely interesting at real work, and the year or two after that is when we find out which ones pencil out without a pilot budget behind them.

Either way, it won't take 40 years to know. Anthropic says it plans periodic updates to its robot exposure index.

I'll be very curious what the next one says.

Sources

• Anthropic, What work can robots do? (Russell Legate-Yang and Maxim Massenkoff, Sept. 30, 2026), plus appendix
• Forbes: Humanoid Robot Shipments Jump 432%, China Has 95% Market Share
• Forbes: Feather Launches $29,990 Humanoid Robot
• Forbes: Agility Robotics Wants To Make 500 Humanoid Robots Safe, Not Just One
• Forbes: First Humanoid Robot Maker Goes Public In US
• Forbes: Figure Is Doubling Humanoid Robot Deliveries Every Month
• Forbes: World's Biggest Humanoid Robot Maker: The Tipping Point Is Near
• Forbes: Inside The European Factory Where Robots Are Building Robots
• Forbes: The 4 Layers Of Brain Behind One Of Europe's Leading Humanoid Robots
• Forbes: Figure's Billion-Dollar Physical AI Bet Delivered A 6X Jump In Robot Chore Success
• HumanoidMegaHub: Boston Dynamics' new Atlas hands
• HumanoidMegaHub: China's humanoid robot schools
• Hexagon: Hexagon and Schaeffler to deploy a fleet of AEON humanoids
• Agility Robotics S-4 registration statement (Digit 5 illustrative pricing)