I’ve paid Anthropic $200 a month for the Claude Max 20x plan since July of last year — the better part of a year now. In that time it has done everything I’ve asked of it, every day, every week, and I have never once hit a wall. For most of those months I carried the same low-grade question: am I overpaying for what I actually use, or quietly getting away with something? After this past week, I don’t wonder anymore.
Back in April, Claude Code started showing subscribers a dollar figure inside the /usage screen — the API-equivalent cost, meaning what a session would have cost at Anthropic’s published API rates if I were paying per token instead of a flat subscription. I’m not billed for it; on the Max plan it’s hypothetical. But it is the same meter that businesses on usage-based plans pay by for real. Glancing at it, my longer days looked like $70 or $80 of API-equivalent work. I knew I was getting my $200 worth. I just didn’t know by how much — and a single session’s number, eyeballed once in a while, is no way to measure anything.
The bill is real for everybody else
For companies paying by that meter, the figure isn’t hypothetical, and lately it has been brutal. Uber rolled Claude Code out to its engineers, leaned on an internal leaderboard to drive adoption, and burned through its entire 2026 AI coding-tools budget in four months. Its COO, Andrew Macdonald, has been candid that the spending is outrunning the payoff: the link between the AI bills and the product actually shipped, he told Fortune, “is not there yet,” and without it “that trade becomes harder to justify.”
Uber is not an outlier. Arvind Jain, CEO of the enterprise AI company Glean, told CNBC that overblown AI budgets are the number-one topic he hears from enterprises: “Companies are telling us that their AI budgets are getting exhausted in one month or two months, and these are annual budgets.” He called the current trajectory “an unsustainable path.” I run on the same engine these companies are metering, and I lean on it just as hard — I just pay for it with one unmoving line item instead of by the token.
So I built a meter
Last weekend I built codecostmeter.com to find out exactly how big my own discount is. You request access with an email, and the site mails back your sign-in details along with a short shell script — and a PowerShell version — to run on whichever machines you code on. I ran it on my two Windows desktops, my Linux laptop, and my MacBook, and started aggregating my usage from that point forward.
The only thing that leaves my machines is the usage math itself: token counts and the dollar estimate, rolled up by day and by model. Not my Claude login, not my Anthropic API key, not a single line of any prompt, file, or answer. There is a key involved — codecostmeter issued me one when I signed up — but all it can do is post my own usage totals to the meter; it is not my Claude key, and it grants no access to Claude or to my Anthropic account. The installer flips on Claude Code’s metrics export and nothing else — the part of its telemetry that could carry prompt text is never switched on. I can export the whole history to CSV whenever I want to slice it myself, and there’s an opt-in public leaderboard if you care to see how your spend stacks up against other people’s. It stays off unless you turn it on.
One honest limit: it only sees Claude Code running in a terminal. Claude in a browser tab, the desktop app, the new Cowork app — none of that rides on this meter.
What a week on the meter looks like
My first full day came back at $119.30 — already above the $70-or-$80 I had been eyeballing. The first full week was $737.25. That covered 3,211,642 input tokens and 6,504,114 output tokens, plus the cache reads and writes layered on top. My heaviest day was a Saturday, at $163.54. And I wasn’t gaming the number or leaving anything churning overnight; it was an ordinary week of work, and I never came close to the session or weekly caps Anthropic puts on the $200 plan.
The discount, in one number
Four weeks like that one come to $2,949 on the meter, and I pay $200 for the same work — a discount of nearly fifteen to one. Set that beside the news above and it lands harder: whole companies are exhausting annual AI budgets in a matter of months, while one flat-rate subscriber pulls close to $3,000 of metered value in a month and hands over $200 for it.
Anthropic plainly knows how lopsided that is. Back in the summer of 2025 it added weekly caps to the consumer plans, aimed at the small slice of subscribers running Claude Code around the clock. I clear $700 in a normal week and still never come near them. For anyone who actually leans on the tool, the flat rate is, for now, an enormous subsidy.
Why the meter is there at all
Now the part I don’t love. I doubt Anthropic put that dollar figure in front of every subscriber out of generosity. It reads more like getting us used to seeing the number before the flat plans give way to metered pricing — the way the enterprise side already has. A subscription that lets one person do the analytical work of a small team for $200 a month, while costing real money in inference, has to close eventually, and the close won’t be gentle for the heaviest users on the consumer plans.
When that day comes, my employer and I will have to rethink how I use Claude and get more deliberate about which sessions earn the spend. Until then, I’m going to enjoy the discount — and now, thanks to my own meter, I know exactly how big it is. If you run Claude Code and have ever squinted at that /usage figure and wondered the same thing I did, point a machine at codecostmeter.com and find out.
This post took $16.44 to produce, by the meter.