The CLAUDE.md length experiment: 301 lines against none

This is the first of four experiments in Reveneau's Claude Code token benchmark. The question it asked is narrow: what does a long CLAUDE.md cost on a task that does not use it? Reveneau is an AI software development consultancy whose code is written by AI, so token use is a running cost of every build, and a file that is loaded on every request is a cost that is easy to stop noticing. Reveneau is independent of Anthropic, and the method is on the method page.

The words used on this page

A token is a piece of text that the model reads or writes; Anthropic's pricing page, read on 4 October 2026, estimates one token at 4 characters of English [4]. The context window is all the text the model reads in one request. Claude Code sends the whole conversation again on every request; the service stores the unchanged start of it, and tokens stored this way are cache write tokens, while tokens read back from that store on a later request are cache read tokens [3]. CLAUDE.md is a file of instructions that Claude Code reads at the start of every session and includes in every request [2]. The effort level is a setting for how much reasoning the model does before each answer; this experiment left it at the default. A hook is a shell command Claude Code runs on its own at a fixed point; none was used here. A non-interactive run is claude -p with a prompt: Claude Code does the task and exits. A turn is one request to the model and its reply, so a run with several tool calls has several turns. A median is the middle value of five sorted runs.

What was compared

Setting A had no CLAUDE.md in the task folder. Setting B had a CLAUDE.md of 301 lines of text and 62,746 characters, arranged as a title and 20 headed sections of 14 numbered project rules each, none of which concerned the task; counting the blank lines between sections, the file is 342 lines long. The file is published as task/CLAUDE-long.md, and the script run.sh copies it into the fresh task folder for setting B only.

Everything else was the same in both settings: the task of one failing test in a 30-line Node.js project, the same prompt, claude-sonnet-4-6 at the default effort level, Claude Code 2.1.118 signed in with a Claude subscription, the flags listed on the method page, and a fresh copy of the folder for each run. Five runs per setting, on 4 October 2026.

The content of the file was unrelated to the task on purpose. The question was the cost of instructions that are loaded on every request and never used. Anthropic's costs page, read on 4 October 2026, describes exactly that situation: a CLAUDE.md with detailed instructions for specific workflows keeps those tokens present even during unrelated work, which is why it suggests moving such instructions into skills that load only when invoked [1]. A file of rules the task did need would have mixed two effects, the cost of the text and any change in how the model worked, and five runs could not have separated them.

The result

Every figure is Reveneau's own measurement from the run of 4 October 2026, read from the json result of each run, and every row is a median of five. The dollar figure is Claude Code's own list-price figure on a subscription account. The raw rows are in runs.json.

Setting	Pass	Cost median (min to max)	Cache read median	Cache write median	Output median	Turns median	Seconds median
No CLAUDE.md	5 of 5	$0.161 ($0.161 to $0.163)	184,772	25,579	659	7	21
301-line CLAUDE.md	5 of 5	$0.239 ($0.221 to $0.256)	256,450	40,066	769	8	17

All ten runs passed. The five no-CLAUDE.md runs cost $0.161 to $0.163, and the five CLAUDE.md runs cost $0.221 to $0.256, so the two ranges do not overlap.

Reveneau's arithmetic

The median cost rose by $0.078, from $0.161 to $0.239. That is 48 percent of the no-CLAUDE.md median. Cache write tokens rose by 14,487 per run, from 25,579 to 40,066. Cache read tokens rose by 71,678, from 184,772 to 256,450. Output tokens rose from 659 to 769, and the median turn count from 7 to 8. All of these are differences between medians of five runs, computed by Reveneau from the figures in the table.

The median wall-clock time (the time from start to finish on an ordinary clock) was lower with the file, 17 seconds against 21 without. Reveneau recorded the times and did not investigate this, so this page gives no reason for it.

What changed, by Anthropic's documentation

Anthropic's prompt caching page, read on 4 October 2026, says the model keeps nothing between requests, so Claude Code re-sends the full context each time: the system prompt (Claude Code's own fixed instructions), the project context, every earlier message and tool result, and the new message. The service matches the start of each request against content it recently processed and reads the matching part from the cache instead of processing it again [3]. The same page places CLAUDE.md in the project context layer, which is loaded when the session starts and is included in every request after that [3]. The memory page says CLAUDE.md files in the working directory and the directories above it are loaded at launch, and that a file over 200 lines uses more context [2].

Reveneau's reading of the two token rows follows from that. The 62,746-character file is written to the cache once per run, which is the rise of 14,487 cache write tokens, and it is read back from the cache on every later request in the run, which is the rise of 71,678 cache read tokens over a median of 8 turns. Anthropic's pricing page lists claude-sonnet-4-6 at $6 per million tokens for a one-hour cache write and $0.30 per million for a cache read, against $3 per million for new input, read on 4 October 2026 [4]. By those list prices a one-hour cache write costs more per token than new input and a cache read costs one tenth of new input, which the pricing page states as a cache hit costing 10 percent of the standard input price [4]. Across a run the file is written once and read on every later request, and it costs something on every one of the 8 turns.

The experiment did not separate the two effects by price, because the medians in the table are medians of different columns across the five runs, and the parts do not add up to the cost median exactly. The dollar difference is reported as measured, and the token differences are reported as measured.

Anthropic's 200-line guidance, and what 301 against none can show

Anthropic's costs page, read on 4 October 2026, says to aim to keep CLAUDE.md under 200 lines by including only essentials [1]. The memory page says the same: target under 200 lines per file, because longer files consume more context and reduce adherence, and move instructions that matter for only part of the codebase into path-scoped rules that load only when Claude works with matching files [2]. It adds that splitting a file into imports helps organisation but does not reduce context, because imported files also load at launch [2].

This experiment measured one file of 301 lines against no file, and nothing in between. It cannot say what a 200-line file costs, or a 50-line file, or whether the cost rises in a straight line with length. Reveneau's measurement agrees with the direction of Anthropic's advice: text in CLAUDE.md that the task does not need is paid for on every request. It does not measure the length at which the cost starts to matter, and Anthropic's 200 is Anthropic's figure, with the reasoning given on its memory page. The sister page on how long a CLAUDE.md should be covers that guidance and what to move out.

What this experiment cannot show

One invented task of one failing test in a 30-line project is one task. The medians are medians of five, and nothing here is a statistic about Claude Code in general.

The task did not need anything in the file. A CLAUDE.md the task does need could change the number of turns, in either direction, and on this task turns drive the cost because every turn reads the growing conversation from the cache again. The run did not measure that case.

The task took a median of 7 or 8 turns in this experiment. A longer session reads its CLAUDE.md on every one of its turns, so the same file would cost more tokens across a 40-turn session than across an 8-turn one; the run did not measure that either, and the page on what the benchmark cannot show lists the experiments that were planned and not run.

The model was claude-sonnet-4-6, as Claude Code 2.1.118 resolved the sonnet alias on that day; the same version resolved opus to claude-opus-4-7 and haiku to claude-haiku-4-5-20251001, and neither was run in this experiment. The dollar figures are Claude Code's own list-price estimates on a subscription account. They do not show how much of a plan's allowance the runs used, which no Claude Code output reports.

What Reveneau recommends

Reveneau's recommendation, drawn from Anthropic's documentation and consistent with this measurement: keep CLAUDE.md under Anthropic's 200 lines, put only what every session needs in it, and move instructions that apply to one workflow into a skill or a path-scoped rule so they load when they are used [1][2]. On this task, 301 lines of unused rules added $0.078 to a $0.161 run by Claude Code's figure. A reader whose CLAUDE.md is long and whose tasks are short can repeat the measurement on their own project with the do-it-yourself page before deciding what to cut. The token usage guide explains why the project context layer is paid for on every request.