Claude Code Quietly Got 5x More Expensive
I started with an easy quota question. Should I get another Claude account or have I been using the ones I have poorly? Then I counted the logs and the answer was much worse.
The price per month remained unchanged. The machine was identical. This was the kind of work I've been doing for weeks. Claude Code, review loops, long running agent jobs, local automation runs. But the amount of Opus work you can see before your account starts to feel empty is nowhere near what it used to be.
So here is the dull figure that is most important to me: visible input and output from Opus. No cache read, only the tokens where Claude read or wrote normal text. 1,442,423 tokens in this week's rolling seven-day window. Same machine, two previous heavy weeks, 8,927,831 and 8,477,395.
So the current week is 16.2% and 17.0% of those old weeks. A sixth maybe. No, that's not "I used it a bit differently."
That is the same subscription looking like it provides dramatically less of the thing I pay for.
If you want to be generous and include cache creation, that's fine. Creating cache is real work. Cache reads are different things so I still leave out cached reads. Current week for that broader non-cached metric was 4,563,761. The old heavy weeks were 22,155,800 and 23,077,312. That means the current week is 20.6% and 19.8% of the old ones.
So the strict count says there's about 83% fewer. The generous count says about 80% less. In either case, the effective price of the useful work seems to be about 5 times higher.
What I can't explain away is the fresh account from June 29, 2026. It was basically full at about 13:44 CEST. 5h:100% and 7d:100% or so. By around 22:45 CEST it showed around 5h:65% and 7d:84%.
In that period there are 0 visible raw Opus rows.
That's not to say a metered event didn't happen somewhere in Anthropic. That shows the user-visible counting didn't explain the meter. That is the problem.
When I did a real benchmark burst later, the logs did reflect that. That run had 1,246,550 input+output, 5,052,266 cache_create, 6,298,816 non_cache_total, and about $108.55 of raw logged cost.
So no, this is not me making out like heavy usage never happened. The logs reflect the heavy usage. The complaint is the meter collapse and the visible accounting doesn't add up.
This also hits hard as Anthropic publicly announced increased Claude usage limits on May 6. Then we had June, full of Claude status noise, like the June 30 Opus 4.8 incident from 14:31 UTC to 15:28 UTC. And the Reddit usage-limits megathread is full of descriptions of sudden drain and mysterious weekly-meter behavior.
I have the receipt, my local logs.
There could be a sensible internal rule. Maybe failed calls count. Maybe retries matter. Maybe Opus 4.8 is differently weighted. Perhaps Claude Code uses up its quota before the local stats file logs a row. But if that's true, Anthropic has to show it.
Developers don't need a fancy progress bar. We need a ledger. Timestamp, model, surface, input, output, cache_create, cache_read, success or failure, retry count, five hour meter delta, seven day meter delta.
That ledger protects Anthropic if the meter is fair. If the meter got harsher, it protects users. From the outside, Claude Code looks like the same monthly subscription was roughly five times more expensive for the work I actually care about.
What I'm building
Delegate tasks. Get software.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.comI respect your privacy. Unsubscribe at any time.