Every week there’s a new headline that says robots are coming for our jobs. This week most of them came from one report.
On September 30, Anthropic published What work can robots do? by Russell Legate-Yang and Maxim Massenkoff. I read the whole thing. The headline numbers are real. So is the part most of the coverage skipped.
I started my career in electrical engineering and test and measurement. The first thing you learn there is that a datasheet has two columns that matter: what the part can do, and what it costs to make it do that in your system. Anthropic’s report is the first robot study I’ve seen that fills in both columns.
The capability column
Here’s what robots can do today, according to the report:
- Robots can perform 74% of physical tasks in the US. That’s 34% of all working hours.
- Most of that is in limited settings. Anthropic sorts tasks by the kind of environment a robot needs. As a share of all tasks, about 23% can be done by robots only in purpose-built environments like an assembly line, 10% in structured workplaces like a warehouse, and just 1% in unstructured places like a city road.
- Add LLMs and the number climbs. About half of work is exposed to LLMs alone. With robots added, it’s 81%.
If you stop reading there, you write the scary headline. A lot of people stopped reading there.
The cost column
Here’s the sentence that matters most in the whole report: robots are cost-competitive with people for just 0.3% of job tasks.
Not 30%. Zero point three.
Anthropic didn’t get that number by looking up the sticker price of a robot arm. For every exposed task, they estimated the full yearly cost of a robot doing the same output a worker produces. That includes the robot, accessories, integration, installation, maintenance, software, energy, oversight, insurance and decommissioning. Fixed costs get spread over the hardware’s life with a cost of capital.
That’s how a test engineer prices a measurement. The instrument is never the expensive part. The fixture, the calibration, the cabling, the operator and the downtime are.
A datasheet for one job
The report walks through packers and packagers, the largest occupation where robots already win on cost. Here’s that example laid out the way I’d want it on a spec sheet:
| Parameter | Value |
|---|---|
| Robots needed | Several: case packing, material moving, inspection, box erecting, labeling, sealing |
| Purchase and install | Over $2 million |
| Workers replaced | About 14 |
| Service life | About 10 years at an 8% cost of capital |
| Robot cost per worker replaced | About $45,000 a year |
| Human cost for the same work | About $49,000 a year |
| Margin | About $2,500 a year |
| Color | Yellow |
Look at the margin. After all that capital, the robots come out ahead by a few thousand dollars per worker per year. That’s the best case in the report, and it’s thin.
Now look at the jobs where robots lose. Welders are the example I’d point an engineer at. Robots with cameras and AI can already weld on their own. But a welder also positions big parts, climbs ladders to reach joints, checks quality and grinds and finishes. Anthropic estimates the set of robots that covers all of that costs about five times what a human welder costs. The weld was never the hard part. Everything around the weld was.
Cleaning is worse. Dishwashers, janitors and cleaners are paid less than welders, and their robot equivalents are still several times more expensive. The report notes cleaning robots can be up to 10 times slower than people.
The 40-year line
The forward-looking number is the one I’d tape to the wall.
Industry and government data suggest robot prices have fallen roughly 3% a year since the 1990s. For robots to be cost-competitive on 10% of work, costs would need to fall about 70%. At 3% a year, that takes around 40 years.
I checked the math. 0.97 raised to the 40th power is about 0.30. So yes, a 70% drop at 3% a year is about four decades. A 20% drop is about seven years, and the report says a 20% drop would make robots cost-competitive for the physical work of 2.8 million workers. That’s real, but it’s 0.8% of all working time in the economy.
Anthropic is careful to say this could move faster. New robots like humanoids, or new manufacturing methods, could push costs down quicker than the historical curve. They also run a fast scenario. I’ll get to that in a later post. But the base case is a long, slow slope, not a cliff.
What a test engineer takes from this
Three things.
Spec at the system level, not the part level. “A robot can do this task” is a part spec. “A robot can do this task at this cost, at this speed, in this environment, with this much supervision” is a system spec. Anthropic did the second one. Most of the coverage quoted the first.
The environment is a spec line. The report’s tiers are basically operating conditions on a datasheet. A part rated for a lab bench isn’t rated for a truck. A robot that works on an assembly line isn’t rated for your kitchen. That’s why only 1% of tasks are doable in unstructured settings.
The cost includes people. Anthropic’s cost estimates often include human supervision, exception handling and repair. Cost parity doesn’t mean the people are gone. It means some of them are doing different work. The report says it directly: cost parity “needn’t imply high unemployment or rapid economic growth.”
Where I land
The capability column is moving. The cost column is moving slower. The gap between them is the most useful thing in this report, and it’s the part that decides when any of this shows up in a real business.
Over the next few posts I’m going to take this report apart one piece at a time: duty cycle and batteries, why hands are the hard part, what the last 20% of work looks like, and why I’d monitor a humanoid the same way I monitor a production service.
Read the report yourself. Then read the cost section twice.