Back to the blog
Articleclaude codetroubleshootingagente de programacaodev tools

Why did my [CC] extra usage credits run out so fast?

MafraNovember 17, 20266 min read

You reloaded extra usage credits in [CC], sent a handful of prompts, and the balance evaporated. The first question is whether the meter is wrong. There are two possible explanations, and the official docs and the bug thread point in opposite directions: it may be real consumption you never saw happen, or it may be a metering and billing fault.

This article shows how to separate the two with the numbers [CC] already gives you, what each number actually means, and what to collect before filing a ticket so the report is not just an opinion.

Why did the credits run out so fast?

Because the cost of a session grows with the context it carries, not with how many questions you type. A session open since morning resends the whole conversation on every message, including the one-line one. A few prompts in a big session cost more than many prompts in a clean one.

The second hypothesis is billing. On 2026-11-05, issue anthropics/claude-code #107558 recorded a $10 reload consumed in roughly 15 minutes of light use, against a history where the same reload lasted half a day or more. The author lists the causes as unverified, which is exactly why the first section of this article matters: without measuring, nobody can say which of the two happened.

What does /usage show, and what does it not show?

/usage opens the usage panel. Under Session you see the token usage of the current session. The official docs are explicit about a detail almost everyone misses: the dollar figure is computed locally, from token counts and the model's list price.

That has two practical consequences. First: if you have a discount, credits or a plan that does not use list price, the gap between the number on screen and the invoice is expected, not a bug. Second: on a Pro, Max or Ultra subscription, the session cost is not your charge, because usage is already included in the subscription. In that case what counts is the usage breakdown and the plan bars shown on the same screen.

What you seeWhat it is good forWatch out
Session cost in /usagefinding which session is expensivecomputed locally at list price
Context size in /contextseeing what fills the windowa big context is resent every turn
Usage breakdown and plan barsthis is the number that counts on a subscriptiononly shown to subscribers
Credit balance and spend in the accountthis is the number that counts on the invoiceit is not inside the session

So /usage is great for comparing sessions against each other and useless for proving the billing is wrong. For that you need the account figure, not the session figure.

How do you find the session that is burning it?

The path is comparison, not guessing. Do it in order:

  1. /context in the suspect session. It shows what fills the context window: files read, tool results, history. If the number is big, that is where the cost lives.
  2. /usage to see that session's consumption and compare it with a session you know was cheap.
  3. /cost for the running total, which is what you paste into the report.
  4. Note the model in use with /model. Changing models changes the price per token, and a bigger model you did not consciously pick is one of the causes listed in the bug report.

If the expensive session has a small context, few turns and no automation running, the consumption hypothesis weakens and the metering one gains ground. If it has a big context and many turns, the opposite.

What burns quota without you noticing?

Anything that fires a request carrying the context along. The cost docs frame this as context management, model selection, extended thinking configuration and preprocessing hooks. In practice, the usual suspects:

SuspectWhy it costsHow to check
A long session never clearedthe whole conversation goes with every message/context and /usage
Reading a large filethe content stays in context and is resent/context
A bigger model than neededhigher price per token/model
Extended thinkingthinking tokens are billed as outputthinking configuration
Subagents and parallel instanceseach one sends its own requestshow many sessions are open
Automation and loopsrequests fire while you are not watchingbackground tasks
MCP servers and hooks injecting contextit enters the prompt on every turnlist of MCP servers, plugins and hooks

The first row is the one that fools people, because it does not feel like consumption: you did nothing, you just left the session open. But every question pays for the accumulated context, and the accumulation grows on its own.

How do you cut consumption right now?

Three commands handle most of it:

  • /clear when the next task is unrelated to the previous one. Clearing costs nothing and is the most efficient cut there is.
  • /compact when you must continue the same task but the context is already large. Summarising keeps the thread and drops the resend.
  • /model to use the model that fits the task, instead of leaving the most expensive one as the default for everything.

For the structural part, the docs point to four levers: context management, model selection, extended thinking configuration and preprocessing hooks. The first two are habits, the last two are configuration. And if the problem is spend across a whole team, the same docs cover setting spend limits for your organisation.

What if the math still does not match?

Then you stop optimising and start reporting. The 2026-11-05 report in issue #107558 is useful as a template because it lists exactly what a maintainer needs to investigate: the output of /cost or /usage, the context size at the time of the incident from /context, the session IDs, the list of active MCP servers, plugins and hooks, and the time window with its timezone.

Collect that before you file. "It burned too fast" is not diagnosable. "In this window, on this model, with this context and this usage, the invoice does not add up" is.

The same report asks for an improvement worth noting: a per-turn breakdown with input, output, cache and origin, plus a spend-rate alert. Until that exists, what you have is the /usage and /cost pair and the comparison between sessions.

How do you keep it from happening again?

Build a baseline. Run a typical task in a clean session, note the consumption in /usage, and use that number as your reference. When a session costs far above the baseline, you know it is context, not price. And when the account spend diverges from what the sessions add up to, you already have the material for the ticket.

Most of this work exists because every token is metered and billed, and because a long session keeps costing while it sits open. In Verboo Code tokens are unlimited: leaving a session open does not become a line on the invoice, so you clear for organisation, not out of fear of the balance.

Enjoyed this article?
Share knowledge with your network.
// Read also

Related articles