### Preflight Checklist - [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code ### What's Wrong? Summary A single Claude Code conversation consumed 26% of my Claude Teams Premium subscription, but third-party token tracking (ccusage) reports only

[BUG] Massive discrepancy between actual token usage and subscription usage metering (26% of Teams Premium for ~$4 of tokens)

Status Fixed / completed

Reported on v2.1.119

Maintainer reply None cached

Activity 6 comments · opened Apr 26, 2026 · closed May 9, 2026

Preflight Checklist

[x] I have searched existing issues and this hasn't been reported yet
[x] This is a single bug report (please file separate reports for different bugs)
[x] I am using the latest version of Claude Code

What's Wrong?

Summary

A single Claude Code conversation consumed 26% of my Claude Teams Premium subscription, but third-party
token tracking (ccusage) reports only ~$4 of actual tokens (180 input, 33K output). There is a ~50-75x
discrepancy between metered subscription usage and actual token consumption.

Environment

Claude Code version: 2.1.119
Model: claude-opus-4-6 / effort level: high
Plan: Claude Teams Premium
OS: macOS Darwin 25.4.0, arm64 (Apple Silicon)
Node.js: v20.19.4

What happened

Over the course of a single conversation (~10 turns), my subscription usage meter increased by 26%. The
conversation involved:

Investigating a cache policy question (spawned one Explore agent)
Analyzing token usage / context sizes (file reads, bash commands)
Optimizing CLAUDE.md files (file writes/edits)
Investigating what's causing the high usage

The actual token usage per ccusage:

Input tokens: ~180
Output tokens: ~33,000
Estimated cost at API rates: ~$4

Yet the subscription meter shows this consumed 26% of the session allocation, according to the /statusline about ~$200-300 worth of usage at whatever internal rate is being applied.

Conversation statistics

Turns: ~10 user messages, ~10 assistant responses
Tools used: Bash, Read, Write, Edit, Glob, Grep, Agent (Explore), ToolSearch
Subagents spawned: 2 (one statusline-setup, one Explore agent)
Files read: ~15 files (CLAUDE.md files, rules files, settings)
Files written/edited: 7 files (CLAUDE.md optimizations)
Total context loaded per turn: ~52 KB of CLAUDE.md + rules files (before optimization), plus system

prompt, MCP tool definitions (~50 deferred tools), 5 enabled plugins, memory files

Context that was loaded

Root CLAUDE.md: 12.4 KB
api/CLAUDE.md: 5.9 KB
app/CLAUDE.md: 8.5 KB
specs/CLAUDE.md: 3.0 KB
7 rules files: 26.3 KB total
Memory files: 3.3 KB
MCP servers: XcodeBuildMCP (30+ tools), ios-simulator-screenshot, Atlassian, GitLab, Linear, Slack
Plugins: android-reverse-engineering, frontend-design, swift-lsp, vercel, claude-mem

Configuration

{
"model": "opus-4.6",
"effortLevel": "high",
"teammateMode": "in-process",
"enabledPlugins": 5,
"hooks": 2 (Notification, Stop — lightweight shell scripts)
}

Key details

This issue started ~2 days ago (around 2026-04-24). No settings changes were made.
The issue reproduces across multiple projects with different context sizes and configurations.
Dormant sessions: ~1,683 total session files across all projects, though dormant sessions should not

consume tokens.

The discrepancy is consistent: early in the conversation, a single Explore agent request was reported

as 1% of subscription. By the end, 26% was consumed for what ccusage measures as $4 of tokens.

What Should Happen?

A conversation with ~33K output tokens on Opus should consume a proportional fraction of the
subscription, roughly consistent with the API cost (~$4). Instead it consumed 26% of the session
allocation.

Error Messages/Logs

Steps to Reproduce

I'm not sure to be honest, as this issue started since a couple days and I checked with a colleague and he's not experiencing the same issue. So it must a be an edge case scenario which is not easily reproducable.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.119

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_

View original on GitHub ↗

6 Comments

0xbrainkid · 4 months ago

The size of the discrepancy is the key signal here. This is not normal "tools add overhead" or "subscription pricing differs from raw API cost" territory — a 50–75x gap between observed token usage and consumed subscription percentage points to a metering/accounting problem, not just an expensive conversation.

A few details stand out:

the raw usage is modest enough to reason about (~33K output, tiny input)
the user already identified one early sign: a single Explore/subagent request showing as ~1%
the behavior appeared recently without config changes and reproduces across projects

That makes me suspicious of one of three classes of failure:

subagent / tool activity is being multiplied or mis-attributed in the subscription meter
some hidden context or replay path is being charged repeatedly per turn
the Teams Premium usage meter is using the wrong internal weighting for this account/configuration

Either way, the UX problem is severe because users cannot sanity-check the meter against any visible work. If ~10 turns and ~$4 worth of tokens can look like 26% burn, people will stop trusting the quota display entirely.

Would be especially useful to compare the per-turn internal usage accounting for the main session vs spawned agents, because the “Explore agent costs 1%” clue sounds like one likely amplification point.

github-actions[bot] · 4 months ago

Found 3 possible duplicate issues:

This issue will be automatically closed as a duplicate in 3 days.

If your issue is a duplicate, please close it and 👍 the existing issue instead
To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

lenkaiser · 4 months ago

@0xbrainkid Yes you're completely right and as a very high load user this reduced my work load to 1/10th while other people (even colleagues) with in the same organization with the same tier don't experience this.

I have to clear one thing up; The Expore that consumed 1% was a relatively simple question:

> Can you check for me if we have setup our cache policy correct?
...
⏺ Explore(Investigate club cache policy)
  ⎿  Done (42 tool uses · 89.9k tokens · 1m 48s)

The 26% was rewriting several CLAUDE.md files and creating a couple rule files:

⏺ Here's the result:

  Before vs After

  ┌──────────────────────────┬─────────┬─────────┬───────────┐
  │           File           │ Before  │  After  │   Saved   │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ Root CLAUDE.md           │ 12.4 KB │ 2.5 KB  │ -80%      │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ api CLAUDE.md        │ 5.9 KB  │ 4.9 KB  │ -17%      │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ app CLAUDE.md        │ 8.5 KB  │ 7.5 KB  │ -12%      │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ api orchestration    │ 8.7 KB  │ 1.9 KB  │ -78%      │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ app orchestration    │ 8.6 KB  │ 2.0 KB  │ -77%      │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ New shared orchestration │ —       │ 3.9 KB  │ (new)     │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ New footguns rules       │ —       │ 2.3 KB  │ (new)     │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ Other rules              │ 7.6 KB  │ 7.6 KB  │ unchanged │
  ├──────────────────────────┼─────────┼─────────┼───────────┤
  │ Total                    │ 51.7 KB │ 34.0 KB │ -34%

Two days ago a request like this would normally consume maybe 1% (or less?) not the 26% it does now. I've been researching what a session defines at Claude and this is not a one to one translation to tokens but also tool calls, inference, context etc etc so I understand it takes more then that. But the difference is just too massive from being able to work and prompt the entire day to hitting session limits in 30 minutes.

I have changed nothing to my setup/plugins or even the way I request claude to execute tasks. I noticed this issue since April 25th in the morning becase I did an equal task I did on Wednesday and hit session limits in 30-ish minutes.

mthormann-dhi · 4 months ago

Same thing going on with me. Started yesterday

lenkaiser · 3 months ago

Maybe a bit premature but I've been working with claude all day today and it seems that due to the recent changes in policies of Anthropic (link) it looks like it behaves like it did 2 weeks ago. The usage is now calculated correctly and it doesn't spend abnormal amounts of usage (token amounts are equal).

I feel that it's not all related to doubling the usage I notice a significant difference in how "expensive" tokens are compared to the session limits. I'm convinced Anthropic was doing some shady stuff but were not publicly admitting to this. For now we can close this ticket but it would've been nice of Anthropic to be MUCH more transparent in this area.

github-actions[bot] · 2 months ago

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.