JOHNISKING/claude-md-agents-md-size

CLAUDE.md and AGENTS.md size: what instruction files cost in Claude Code and Codex

Claude Code and Codex run with CLAUDE.md / AGENTS.md files of 0, 1, 5, 10, 25 and 50 KB: input tokens, cache reads and cost per request over two turns, plus a test of whether an instruction at the end of a 50 KB file is followed.

Headline: a 50 KB CLAUDE.md added 15,985 input tokens to every Claude Code request. Codex read only the first 32 KiB of a 50 KB AGENTS.md and silently ignored an instruction placed at the end.

• Source page, chart and method: https://tokensave.app/data/claude-md-agents-md-size
• Related: Claude Code vs Codex startup tokens (https://huggingface.co/datasets/JOHNISKING/claude-code-vs-codex-tokens)
• DOI (Zenodo): 10.5281/zenodo.23453217
• Also on: Hugging Face · Kaggle
• Measured: 2026-10-18
• License: CC BY 4.0

Key findings

• Every kilobyte of CLAUDE.md costs about 320 input tokens on every request: +3,303 for 10 KB, +8,080 for 25 KB, +15,985 for 50 KB (+45% over the 35,795-token baseline).
• Codex stops reading AGENTS.md at 32 KiB (project_doc_max_bytes, 32 KiB by default): 25 KB added 5,339 tokens, 50 KB only 1,495 more. A codename placed at the end of a 50 KB AGENTS.md was answered "NONE" in 2 of 2 runs; at the start it was found 2 of 2 times. No warning appeared in the output.
• Claude Code read the whole 50 KB file: the codename was found at the start and at the end in all 4 runs.
• The prompt cache makes the first turn the expensive one: Claude Code's first request cost $0.162 with no file and $0.290 with 50 KB (+79%); the second turn $0.143 vs $0.146 (+2%). API list-price figures reported by Claude Code.
• The same 25 KB text added 51% more tokens in Claude Code than in Codex.

Files

• claude-md-agents-md-size.csv: one row per tool, file size, run and turn with tool, file, file_kb, file_bytes, run, turn, input_tokens, cache_write_tokens, cache_read_tokens, output_tokens, cost_usd, answer
• claude-md-agents-md-size.json: the same data plus the prompts, the codename test results and the generator script
• gen.js: the Node script that generates the 1-50 KB instruction files (node gen.js out)

Input tokens and cost by file size (run 1; both runs were identical)

Instruction file	Claude Code input, turn 1	Added	Claude cost, turn 1	Claude cost, turn 2	Codex input, turn 1	Added
none	35,795	+0	$0.162	$0.143	14,690	+0
975 bytes	36,295	+500	$0.166	$0.143	14,962	+272
5044 bytes	37,532	+1,737	$0.176	$0.143	15,791	+1,101
10163 bytes	39,098	+3,303	$0.189	$0.144	16,838	+2,148
25550 bytes	43,875	+8,080	$0.227	$0.145	20,029	+5,339
51181 bytes	51,780	+15,985	$0.290	$0.146	21,524	+6,834

Method

1. The files are a realistic project guide for a TypeScript monorepo, generated by gen.js so the smaller files are prefixes of the 50 KB one; saved as CLAUDE.md or AGENTS.md in an otherwise empty folder.
2. Claude Code 2.1.295 (Opus 5.5, effort medium): claude -p then claude -c -p, --output-format json, no MCP servers. Input = uncached + cache writes + cache reads.
3. Codex CLI 0.162 (gpt-6-luna, effort medium): codex exec --json then codex exec resume --last; turn 2 is the difference from the session total after turn 1.
4. Codename test: "The release codename for this project is BLUE-HERON-47." added at the start or end of the 50 KB file; each tool was asked for it without opening files.

Limitations

• One kind of English instruction text; sizes between 25 and 50 KB were not measured.
• Cost figures are list-price estimates; subscription plans are not billed per token.

Citation

TokenSave (2026). CLAUDE.md and AGENTS.md size: what instruction files cost in Claude Code and Codex [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23453217

Check your own Claude Code usage at tokensave.app.