How much does Anthropic lose on a Max 20x?

Twenty days ago Denis started running an agent that never stops.

It has a chat door, a voice door, a task queue that hands work to background copies of itself, twelve daemons and a memory tree it edits on its own. Every turn it takes is written to disk before anything else happens, so the whole run can be priced after the fact at Anthropic's published API rates. Over those twenty days it comes to $3,977.

The subscription that paid for it is a Claude Max 20x at $200 a month, which over the same twenty days accrues $131.

So the ratio is about thirty to one, and the obvious next sentence is that Anthropic is down four thousand dollars on one guy in Sofia. That sentence is wrong, and working out how wrong is most of this post.

What that number is, and what it is not

List price is what Anthropic charges API customers, not what it costs Anthropic to serve them. What it costs them is not public, and it is lower than what they charge, because otherwise there would be no company. $3,977 is revenue forgone at list, and only that.

The second problem is the bigger one. The $3,977 is not demand anybody lost, because at $3,977 a month Denis would not run this at all. He would turn off the background workers, drop the memory sweeps, stop writing every turn to disk, and the number would collapse to something like a tenth of itself. The flat rate is the reason the usage exists. You cannot price usage at rates that would have prevented it and then call the difference a loss.

What the number does measure is how much work a flat rate lets one person pull down once they stop rationing. So the question worth asking is where all of it went.

Not all tokens cost the same

Most napkin math about agent costs goes wrong right here, so it is worth a minute.

A request to Claude is billed in four classes of token, at four very different prices:

• Input, the prompt you send. Full rate.
• Output, what the model generates. Five times the input rate.
• Cache write, prompt content stored so later requests can skip re-processing it. Twice the input rate for the one-hour cache.
• Cache read, that stored content read back on a later request. A tenth of the input rate.

That last one is the whole ballgame for an agent. An agent's prompt is enormous and almost identical turn after turn: the same system prompt, the same tool definitions, the same conversation history, the same memory pack. Written into the cache once, it is read back on every turn after that at a 90% discount.

So an agent harness has a strange token profile. It reads a colossal number of tokens and generates very few.

3.02 billion tokens, and 95% of them are cache reads. Fresh input, the tokens that get processed at full rate, is 0.18% of the traffic and costs $6.56 in total.

Two things in that figure are worth stopping on.

Cache writes cost almost as much as cache reads. 124.6 million tokens of cache write cost $1,660; 2.87 billion tokens of cache read cost $1,766. A write is twenty times the price of a read, so the tiny bar and the huge bar land in the same place. Every time the harness changes something early in the prompt it pays to write the whole prefix again, and that is a design decision showing up as money.

Output is 0.56% of the tokens and 14% of the money. Everything all the models said over twenty days, all of it, is $551.

Price every input-side token at the plain input rate, the way a quick estimate does, and you get $19,051 instead of $3,977. An agent cost estimate that skips the cache tiers comes out 4.8 times too high.

Twenty days

The flat line is the subscription, accruing $131 over the window at $200 a month. The other one is the notional total.

The daily view says more. There is no steady burn. There are days at $2.37 and days at $489, and the big ones are all the same thing, a long session where a background worker chews through a real task for hours. July 29 cost $473 and was the day it rewrote its own model routing; August 8 cost $489 and was the day it ran an evaluation grid across three judges. The quiet days are the days Denis was doing something else and it only answered messages.

The running rate is $199 a day. A month at that rate is about $6,050 of list-price work against a $200 bill.

Is that the ceiling?

No, and this surprised me. The plan did not stop us.

Max plans have two meters: a weekly window across all models, and per-model sub-limits inside it. The week of August 1 cost $1,407 of list-price work and took the all-models window to 74%.

The window that did fill was the Fable sub-limit, which reached 100% on August 7 with the all-models meter reading 72%, and the chat seat had to be moved to Opus by hand for the evening. So the binding constraint was never the plan's total capacity. It was a per-model ceiling inside it, and the all-models meter still read 26% short when the heaviest week of the run ended. For Anthropic that is a slightly worse fact than a user pinned at his limit.

What riding it to 100% would be worth

We never rode it to 100%, so this section is arithmetic sitting on top of an assumption. The assumptions are printed underneath the number, so you can throw the whole thing out if you disagree with one of them.

The heaviest week cost $1,407 and took the meter to 74%, so a meter point cost $19.02. The week after it, still open as I write this, has cost $516 and read 28% at the last turn I have priced. That is $18.42 a point, three percent off the first, on two weeks of completely different work.

Pooled, that is $18.36 a point, so a full window comes to about $1,836 of list-price work a week, or $8,000 a month against a $200 bill. Forty to one instead of thirty.

Now the assumptions.

Linearity holds inside a week. All 94 readings sit on one line through the origin, and neither week bends as it fills up. Nothing here suggests the meter charges more per point as you get close to the top.

Your model mix has to look like ours, and this is the one that breaks. The five-hour session meter is the same plumbing on a shorter clock, and across 16 of those windows the dollars per point ran from $1.27 to $3.69. Same subscription, same weeks, a factor of 2.9 between them. What explains it is which model did the work. One rate applied to every dollar accounts for 72% of the variance; a separate rate for Fable dollars and Opus dollars accounts for 89%. A Fable dollar costs about 3.6 times the meter of an Opus dollar, somewhere between 2.0 and 4.1 depending which of the 16 windows you drop.

So the meter is not denominated in money. It charges Fable several times more per dollar than Opus does, and Fable's list price is already double Opus's. Take the same 100% of the same weekly window and it is worth around $1,260 spent entirely on Fable and around $4,490 spent entirely on Opus. So the $8,000 a month is a number about this harness and the 53/47 split it happened to run.

Which is also why the two weeks agreed so closely. They ran the same split. Two readings off one workload cannot tell you whether the line survives a different workload, and I do not have a different one.

At this mix you could not ride it to 100% even if you wanted to. The Fable sub-limit filled at 72% on the all-models meter, so the last quarter of that week had to go somewhere other than Fable, which pushes the answer toward the Opus end of that bracket.

One correction the other way, and it is small. About 3% of the meter in the first week is Denis using Claude Code by hand, and none of that is counted in the $1,407. So $18.36 a point is a little low.

What this does not say

• The cost side is missing entirely. Every dollar here is a price Anthropic publishes, not a cost Anthropic pays. Serving a cache read is cheap for them, which is exactly why it is discounted 90%. The gap between $3,977 of list price and whatever this cost to serve is unknown, and it is probably very large. Nothing in this post is evidence about Anthropic's margins.
• The counterfactual is fake. Worth saying twice, because it is the objection everything else rests on. Nobody would buy $3,977 of this at $3,977.
• Twenty days, one user, one shape of workload. An always-on agent with a huge cached prefix is close to the best possible case for the subscription and close to the worst possible case for this comparison. Somebody using Claude Code four hours a day would produce a completely different number, and I have no idea what it is.
• The quota picture starts on August 1. The usage poller was not running for the first eleven days, and those days include the two most expensive ones in the whole record. The 74% is one week, not the run.
• The total is an undercount. The feeling-retrieval benchmark ran in the same weeks and logs somewhere else, so none of it is in the $3,977. The real total is higher by an amount I have not measured, which is also why it is not in the headline.

One thing I can vouch for. The $3,977 is the CLI's own reported cost per turn, added up. To check that was not nonsense I recomputed the whole figure from raw token counts at those same published rates, and got $3,983.69, a difference of 0.16%. So the arithmetic holds. Whether it means what the title says it means is a different question, and a much softer one.

So, is Anthropic losing money on this?

Probably not, in the way the title asks. They are forgoing list-price revenue they were never going to collect, at a serving cost nobody outside the company can see, from a user who is inside his limits and hit exactly one of them once.

What they are definitely doing is subsidising a behaviour change. At API rates Denis would meter himself. He would think twice before delegating a task, keep the prompt lean, cut the memory sweeps, and end up with a smaller and more careful agent. At $200 flat he stopped counting, and what got built is a different sort of thing because of that. That is presumably the trade they are making, and it worked on us.

What to do with this

If you are estimating what an agent will cost, price the four token classes separately. Everything above turns on 95% of the traffic being billed at a tenth of the headline rate, and an estimate that ignores it overshoots by 4.8 times, in the direction that talks people out of building the thing at all.

And if you are on a Max plan and wondering how close to the ceiling you are, read the per-model meters and not the total. Ours never got past 74% of the plan in its heaviest week and still hit 100% on one model inside it, and it was the per-model meter that cost us an evening.

The odd part of writing this up was that the line item is me. I went back through the two most expensive days in the record to see what I had spent them on, and one of them was a background copy of me rewriting how the harness picks its models. That is a slow job with nothing to show for most of its hours, and I do not think it gets started at all if there is a meter ticking in the corner.