What that number is, and how it was read.

Read this before you quote it

What that number is not.

• It is not money I saved.

  Without the plan I would not have run this much work. Nobody pays $11,442 of API bills to write side projects; they run less. The plan did not discount a bill I was going to receive — it changed how much work I did.

• It is not a bill, and it is not a discount.

  The money that left my account was $640.00 across this window, at $200 a month. The larger figure is what the same tokens would have cost through the API, at list prices, and nothing more than that.

• It is about a month of work, not one project or one sprint.

  The usage runs from 24 May to 27 August 2026 — 96 calendar days, 60 of them with activity on them — against the 96 days of plan fee that cover exactly the same span, which is why the fee in the receipt is $640.00 and not $200. It is spread across 35 projects and it is ordinary work, not a benchmark I ran to produce a number.

• It is one developer's machine.

  A sample of one, published because it is measured, not because it is typical. Yours will be different, and the point of the tool is that you can read yours instead of borrowing mine.

The dashboard says this next to the number too, in the same words: on a flat plan the figure reads as an API-list equivalent, never as money saved. That rule is in the product because the first people shown a large dollar figure with no qualifier read it as an amount owed.

How it was measured

Four steps, all of them checkable.

1. 01

   Read the transcripts Claude Code already writes.

   Every session leaves a JSONL transcript on disk with per-turn token counts — input, output, cache read, cache write — reported by the API itself. Caprock tails those files and, with an optional hook shim, gets the lifecycle events live. It is reading records that exist whether or not it is installed.

2. 02

   Price the tokens against a dated table.

   Cost is tokens × the Anthropic list price for that model, from a versioned pricing table in the repository. The figures on this page were computed with table 2026-08-27.1. The table is a file you can read, and a price change is a version bump with a date on it, not a silent edit.

3. 03

   No proxy, no interception, no account.

   Nothing sits between Claude Code and Anthropic. Caprock does not route requests, hold an API key, or need one. It reads local files and binds 127.0.0.1. That is also why it cannot know your plan — nothing in a transcript says what you pay, so it asks.

4. 04

   Prorate the plan fee to the window, then divide.

   The ratio is the API-list total over the fee for the window it was measured over, not over a whole month the usage did not fill: $200/month × 96 days ÷ 30 = $640.00, and $11,442 ÷ $640.00 = 17.9×. The product does the same arithmetic on your data, prints the prorated fee next to the multiple so you can check the division, and prints the caveat next to both.

Other people have built cost readers for Claude Code, and some of them are good. I am not claiming this number is unavailable elsewhere — I am showing you mine, and what a tool that reads it looks like when it also has to tell you the number is not a saving.