Where did 16 billion Claude Code tokens actually go?
TL;DR I ran one read-only script over 39 days of my own Claude Code logs. 97 of every 100 tokens were the model rereading context it had already seen, and that rereading was 55% of the list-price estimate. Another surp
TL;DR
I ran one read-only script over 39 days of my own Claude Code logs. 97 of every 100 tokens were the model rereading context it had already seen, and that rereading was 55% of the list-price estimate. Another surprise: 46% of all tokens came from helper agents whose numbers I had never looked at.
The same report showed Haiku doing 29% of my calls for 1.3% of the estimate, a median session peak of 271,000 tokens of context, and 183 of my 253 installed skills never called.
The script is the
usage-dashboardskill, free in the Second Brain Starter. One command, no network, no prompt text printed.
I write Code Meet AI, the newsletter where I document how I actually use AI to ship real products, receipts included.
You are probably guessing where your Claude Code tokens go. I was. I assumed the expensive part was the code it writes. Then I counted.
Between 13 October and 20 November, my machine logged 270 Claude Code sessions, 75,055 model calls and 16.6 billion tokens. Priced at Anthropic's API list prices, that is about $4,562, a list-price estimate and not a bill. The script says so itself on its last line, which is more honesty than most dashboards manage.
97 of every 100 tokens were rereading
Every call sends the conversation back to the model. Most of it comes back from a prompt cache instead of being processed fresh, and each transcript records those tokens in their own field. The dashboard calls them a cache read.
Over the 39 days, cache reads took 16.04 billion tokens, or 96.8%. Cache writes, the context stored for reuse, took 2.8%. Output, the code and text the model wrote, took 0.4%.
So the part I worried about, the writing, was under half a percent of the tokens. Cache reads are cheap per token. In the price file the script ships with, Opus 5.5 costs $0.20 per million tokens for a cache read and $4.00 for fresh input. Cheap times 16 billion is still money. With that same price file I multiplied the cache-read tokens out per model: about $2,510 of the $4,562, or 55%, was rereading.
The cost is the size of the context, paid again on every call.
Almost half the tokens came from helpers I never watched
Claude Code can start sub-agents: helper runs with their own context that do one job and report back. Each one writes its own transcript in a subagents folder, and the dashboard adds it into the session that started it.
On my machine, 45.8% of all tokens came from sub-agents. I run many of them on purpose: fresh reviewers, scorers, research tabs. Still, I had never seen the number, and each helper builds its own context from zero.
The median session peaked at 271,000 tokens of context
Peak context is the biggest single call in a session's main thread: input plus cache read plus cache write. It tells you how full the window got before you cleared or compacted it.
Across the 268 sessions that had a peak, the median was 271,434 tokens. 41 sessions went past 500,000, 14 went past 800,000, and the biggest reached 968,405.
My vault already has a rule to roll a long session over before it gets this big. The numbers say the rule and I are not always in the same session.
Haiku did 29% of the calls for 1.3% of the estimate
The by-model table was the most useful part for me:
| Model | Share of calls | List-price estimate |
|---|---|---|
| Sonnet 5.5 | 37.5% | $2,304 |
| Opus 5.5 | 30.1% | $2,200 |
| Haiku 5.5 | 29.3% | $58 |
Haiku ran 21,975 calls for $58, 1.3% of the estimate. Opus ran 22,627 calls for $2,200. That does not mean every Opus call should have been Haiku: my planning and long writing run on Opus by choice. It does mean a task I move to Haiku is close to free at list price, and the table shows me which bucket each kind of work landed in.
One limit, stated plainly. 2,276 calls, 3% of the total, ran on four model ids that are missing from the price file, which is dated 6 October. The script marks them "no price" and leaves them out of the money. It never counts them as zero.
183 of my 253 skills were never called
The skills section counts how often the Skill tool loaded each skill. I compared those names with the skills installed on my machine: of 253, 70 were called at least once in 39 days and 183 never were.
The dashboard cannot list the 183 for you, because it only sees what ran. You get them by comparing its list with your skills folder, which takes one more minute.
Before you delete anything: a skill the model opened as a plain file does not count as a use, and some skills exist for the one bad day a month. The script also does not measure what an unused skill costs you. I treat the 183 as a review list and start with the ones I cannot explain.
How to run it on your own machine
The skill ships in the free Second Brain Starter, and its installer puts it in .claude/skills and .agents/skills. From the root of a workspace set up with it:
node .claude/skills/usage-dashboard/usage-dashboard.mjs
That prints a header, the totals, the by-model table, the top 20 sessions, then skills and commands. It needs Node 22 or newer.
node .claude/skills/usage-dashboard/usage-dashboard.mjs --since 2026-11-01
node .claude/skills/usage-dashboard/usage-dashboard.mjs --top 0 --json
The first line keeps only sessions that started on or after a date. The second prints every session as one JSON object, which is the easiest form to add up yourself.
On my laptop it read 1,713 transcript files in 12.8 seconds. What it leaves alone matters as much:
It opens no network connection. Its only imports are Node's file, path, OS and URL modules.
It writes no file.
It prints no prompt, answer, tool input or session title. You get counts, model and skill names, session ids and project folder names.
That last point has a catch, and the skill states it itself: project folder names come from your folder paths, so they can show a client name. If you paste the output into an AI chat, those names go to that provider. Share the totals and keep the session table to yourself.
Start with three numbers from your own run
Run the command once, with no flags.
Read three numbers: your cache-read share, your sub-agent share and your median peak context.
Pick one task that runs on your most expensive model and try it on a cheaper one this week.
The skill is free and MIT licensed in the Second Brain Starter. Run it once on your own logs.
When you do, what share of your tokens turns out to be cache reads? Tell me in the comments.
I write Code Meet AI, the newsletter where I document how I actually use AI to ship real products, receipts included.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.