I measured where Claude Code actually spends tokens: 96.8% is re-reading history, my typing was 0.01%

Claude Code writes a log of every session to ~/.claude/projects/. Each line is JSON, and message.usage holds the token counts the API actually returned.

I aggregated 32 sessions. These are not estimates.

96.8% is re-reading conversation history

Type	Tokens	Share
Re-reading history (cache_read)	1,370.9M	96.8%
Cache writes	37.7M	2.7%
Model output	6.7M	0.5%
What I typed in	0.2M	0.01%
Total	1,415.5M	100%

My own typing is 0.01%. Shortening prompts optimizes one ten-thousandth of the bill.

The mechanism: the model re-reads the whole conversation every turn. So once a large return enters the context, you keep paying for it on every following turn. Drop a 10k-char return in, talk for 20 more turns, and you paid for 200k chars.

88.8% of that history is tool return values

Not prose. Not your instructions. Tool output.

It is not call frequency. It is return size

Tool	Share	Calls	Per call
Full page fetch (browser)	63.3%	417	33,645 chars
Reading a whole file	13.0%	258	11,132 chars
Screenshot	4.0%	3	296,827 chars
Bash	3.7%	967	839 chars
Extracting only the values with JS	0.4%	164	504 chars

Look at the last two rows.

967 Bash calls cost less than 3 screenshots. One screenshot is ~354 Bash calls.

And in the same browser task, fetching the whole page costs 33,645 chars while extracting just the values you need costs 504. A 67x difference.

Three things that actually move the number

1. Never read a whole file. 11,132 chars per read. Grep for the line numbers first, then read that region. Read the whole thing exactly once, when you first need the structure.
2. Do not take screenshots out of habit. 296,827 chars each. If the state is readable as text, read it as text.
3. Never fetch a whole web page. Extract the handful of values you need. 33,645 becomes 504.

Conversely: stop rationing shell commands. 967 calls came to 3.7%.

Caveats

• This sample skews toward web research, so browser calls rank first. If you mostly write code, file reads will lead. Measure your own logs.
• cache_read is billed at a lower rate than fresh input, so token share is not cost share.
• Whether weekly usage limits deplete in proportion to this split is unverified; the formula is not published.
• Sample: 32 sessions, 2,478 tool returns.

Measure your own

~/.claude/projects/

One folder per project, .jsonl inside. Each line is JSON: message.usage for tokens, tool_result for the returns.

Take the method, not my numbers. Your top consumer will differ, and that is the one worth cutting.