Claude Code: why your quota melts so fast, and how to finally see it

You ask a completely harmless question. Three lines. And you see your gauge drop as if you'd just asked it to rewrite the Linux kernel.

It's been annoying me for months. Not the price, the mystery: impossible to know what just cost so much. Well, for a week now, it tells you.

You, watching your quota disappear without knowing what for

Versions 2.1.257 to 2.1.267 came out between September 1 and 9, and there are four things in them that talk about money. Not shiny new features for demos. Meters.

The thing that costs a lot is rereading everything

First, you need to understand where the cash goes, otherwise the rest is useless.

When you talk with an assistant, it remembers nothing. Nothing at all. Every time you type a sentence, it rereads the entire conversation from the beginning: your instructions, the open files, the previous twenty exchanges. And you pay for that rereading. Every message.

So they came up with a trick. As long as the beginning of the conversation doesn't change, it keeps a copy of it ready to use and reuses it instead of rewriting everything. This copy costs about four times less than a complete reread. That's what they call the cache, and it's what lets you work for five hours without blowing up your quota.

Except it's fragile. Very fragile. All it takes is for one thing to change right at the beginning for the copy to be thrown away and for everything to start from zero again.

One place setting moved, and the server resets all fifteen tables

A tool added along the way. An instruction changed. A configuration file you touched without thinking about it. Or just you, off to eat and back forty minutes later: the copy has a lifespan, and it expires on its own.

And until this month, when that happened, you didn't know it. You just saw that it had cost a lot. Great, figure it out yourself.

Now it tells you why it reread everything

It happened on September 3 in 2.1.260, then was completed on the 9th in 2.1.267: the /cost command and the status line now display the probable cause of the cache break. The official text talks about “modified tool definitions or system instruction, inactivity beyond the lifespan”.

It looks like nothing much, written that way. Yet it's the most useful change of the month, because it turns an inevitability into something fixable. “Your tool list changed during the session”, that can be fixed. “You were away too long”, that can be anticipated.

In my case, the real culprit was the first one. I was adding a tool server in the middle of a session, nice and easy, telling myself it wasn't doing any harm. It threw away the cache. Every time.

The skills that sleep and that you still pay for

This one hurt me.

If you've made skills, those little files that teach your assistant a particular know-how, they're loaded at startup. All of them. Even the ones you haven't used in three weeks. And each one takes up space in the conversation, space you pay for with every message.

2.1.261, released on September 4, adds /skill-doctor. It shows you which loaded skills are useless and what they cost in context, so you can prune them.

Three screwdrivers worn down to the handle, and twenty-eight tools collecting dust since the inauguration

It's exactly the workshop's tool wall. You hung forty things on it, you always grab the same three, and you pay the rent for the whole wall.

Putting a ceiling on it, for real

Two additions from September 9, in 2.1.267.

The first is called maxEffortLevel. Effort is the time the assistant spends thinking before answering: the deeper it digs, the better it is, the more expensive it is. This setting caps that level everywhere, including when you go through Amazon, Google or Microsoft platforms. You can still lower it manually, but you can no longer go above it by accident.

The setting that was missing: the one that stops you going up, not the one that lets you go down

The second is a spending limit bar in /usage. It only concerns teams that go through a corporate gateway, so not you if you're on your own with your subscription, but the principle is right: see the limit while you're working, not afterwards.

One remark anyway, and this one's for Anthropic. They cap effort, display spending, diagnose the cache. All of that is happening at the same time, in September. Translation: people were spending without understanding, and enough for it to get flagged.

The thing that has nothing to do with it but justifies the update

In 2.1.265 from September 8, there's a line that has no business in an article about money, but I can't let it pass.

Permission rules whose path contained parentheses were simply and completely ignored. Result: a folder that you had declared read-only remained writable. You thought you'd closed the door: it was wide open.

Who does it affect? Everyone who keeps their projects in a folder with parentheses in its name. On Windows, Program Files (x86), just to name that one. A month ago, I wrote here that Claude Code had stopped asking for permission at every move, and I was already saying that the scope deserved a closer look. This is exactly the kind of hole I was talking about.

Concretely, what you type this morning

Three actions, ten minutes, and you'll finally know where your money is going.

Update first. I typed claude --version while writing these lines: 2.1.266. One version behind yesterday's, and I hadn't even noticed. With six versions in five days at times, no one knows what version they're running on anymore, me included.

claude update

Then, start a normal session, work for twenty minutes, then look at this:

/cost
/skill-doctor

The first will tell you whether your cache is holding up or falling over, and why. The second will tell you what you're carrying around for nothing.

And if you discover that you're paying three times a day for a complete reread because a tool server gets added along the way, declare it at startup and that's sorted. It's this stupidly simple.

Frankly, this month, Anthropic hasn't released a feature that makes you dream. They've released counters, ceilings and a diagnostic. It's less marketable than a new model, and it saved me more time than the last six.

Come on, go type /skill-doctor. And get ready to discover that half your skills have been asleep since August!

Sources

• Claude Code: changelog, the official source, with versions 2.1.257 from September 1 to 2.1.267 from September 9, 2026, the details of maxEffortLevel, the spending limit bar, the cause of cache misses and the fix for permission rules containing parentheses
• Claude Code: cost tracking, what /cost measures and how to read the status line