Anthropic got the math wrong. And the margin is going to whoever notices.
A partner company let a Claude Opus system run for 43 days straight. Twenty-four hours a day. 72 billion tokens. 1,253 hours of active sessions.
At API rates, that workload is worth 180,000 euros. An equivalent human team, pure manpower, would cost 112,000.
So the story that AI is cheaper is a lie? Yes, if you stop at the price. Most people stop there, shrug, and go back to complaining about the Max plans on the forum.
The mistake is reading the wrong line on the balance sheet. Price is not the metric. Output is. You write 500 words an hour, the system writes twenty-one thousand. You read one document, it digests fifty. Forty-two times more. Replicating those 43 days by hand would take 250,000 person-hours, roughly 23 million euros in salaries. A 200x output multiplier.
This is where the story stops being about cost and starts being about structural replacement. AI is not replacing you because it's cheaper. It's replacing you because it produces volumes a human, physically, cannot reach. Everything else is conference rhetoric.
Two economies, same silicon
Today the market is split in two. On one side, those who pay retail. Pay-as-you-go API, Copilot Pro, getting little for a lot. On the other side, those building infrastructure on top of flat subscriptions, the Claude Max plans, getting two orders of magnitude more output. Same models, same silicon, two different economies.
Anthropic has underpriced Max by a huge factor. This is a window, not a permanent gift, and it lasts maybe 12 to 24 months. Then the rules will change, as they always have for every underpriced infrastructure of recent history, from Uber to cloud credits.
The pricing window you are looking at is the calculation error of a 380 billion dollar giant.
Three practical consequences
If you are a worker: knowing how to use ChatGPT is not enough anymore. Your real competitor is not the colleague at the next desk. It is whoever is industrializing these systems while you wait for the company training course.
If you are a company: stop selling AI as "headcount savings." That is a poor framing for a productivity multiplier you will not see again in your career. Treat it as a cost cut and you miss the strategic train.
If you are a builder: do not settle for what is served to you. Every day this window stays open counts.
P.S. Yes, the partner company accounted for cache reads. Without them, the spend would have been 790,000 euros. If you are paying as you go and you are not calculating cache reads, you are probably burning money you do not know you are burning. Ask your model how it works, before the next invoice.
— Pan