Because every turn sends the whole conversation again. A long session is not expensive once; it is expensive on every single turn after it gets long.
Every turn pays for everything before it
Claude Code has no memory between turns. Each time it replies, it is sent the full session so far: the instructions, the tool list, every file it read, every command output, and your messages. A session that has grown to 300,000 tokens sends 300,000 tokens to answer "ok, run the tests".
That is why usage climbs faster the longer you work in one session. The work per turn stays the same. The amount re-read per turn keeps growing.
The three things that use your tokens
| Cause | What it looks like | Fix |
|---|---|---|
| History | A session that has been open for hours and carries hundreds of thousands of tokens of old output. | Compact it, or start a fresh session for the next task. |
| Overhead before any work | The very first turn is already large, because every connected tool and plugin loads its full description. | Turn off the connectors and plugins this session does not need. |
| Subagents | One request fans out into many helper agents, each with its own context. | Use fewer, narrower helpers, and only for work that needs them. |
How to reduce token usage
- Start a new session for each new task. A fresh session carries no history.
- Compact a long session you need to keep, with
/compact. - Turn off connectors and plugins a project does not use. Their tool lists are re-read every turn.
- Use subagents for wide, independent work only, and give each a narrow job.
What this looked like on a real machine
On the author's Mac, two ordinary sessions each started at about 126,000 tokens before any work was done. That overhead was the tool lists of connected services and plugins. One of them grew to 300,000 tokens per turn within an hour, which made each turn cost about 2.4 times what it would in a fresh session.
Across one week, daily totals ranged from 168 million to just over 1 billion tokens. On the heaviest subagent day, 61 percent of all turns were made by subagents, not by the session the person was typing in.
The figures on this page come from real session logs on the author's Mac, read with Cost Doctor. Your numbers will differ; the pattern usually does not.
How to find out which one is yours
A usage meter tells you how much of your limit is gone. It does not tell you which session used it or why. For that you need the per-session numbers: where each session started, how big it is now, and how much of it was subagents. Claude Code writes all of this to log files on your Mac. See which session is using my limit for how to read them.
See it on your own sessions
Cost Doctor is a Mac menu bar app that reads the logs Claude Code and Codex already write. It names the session that is using your tokens, says why, and turns your screen red when a check fails while an agent sits waiting. Nothing leaves your Mac.
Download the free trialHow it works14 days free, no account. Then $19 once, plus tax where it applies. For Apple silicon and Intel Macs on macOS 13 or later.