The docs say ~100 tokens a skill. I measured 34 to 96.

If you are about to install a pack of Agent Skills, the question is what it costs you before you use any of it. Anthropic answers that, in a table on the Agent Skills page:

Level	When loaded	Token cost	Content
Level 1: Metadata	Always (at startup)	~100 tokens per Skill	name and description from YAML frontmatter

That is the first of three rows; the other two say a triggered skill body costs under 5k tokens and bundled resources cost nothing until read. Only Level 1 is charged whether or not you use the skill, so it is the one this post is about. The emphasis in the cell above is mine.

That is a clear, useful, specific number, and it is the only per-skill figure a user is likely to find. This site has installed six trending skill packs and measured them. Here is how the documented figure did.

What it costs, measured

Pack	Skills	Per skill	Documented	Over-predicts by
addyosmani/agent-skills	24	95.9	~100	4%
google/skills	111	81.3	~100	23%
mattpocock/skills	25	34.1	~100	193%

Every pack where the measurement is reliable came in under the documented figure, and the size of the gap ranges from negligible to nearly three-fold.

Concretely: install google/skills and the documented figure tells you to expect 11,100 tokens on every session. It costs 9,020. Install Matt Pocock’s 25 skills and the figure says 2,500. It costs 852.

One caveat on the Google figure, which the source corpus carries: google/skills ships no plugin manifest, so “installing it” has no single definition. The 9,020 prices copying all 111 skills, and they are not one thing — 96 are Cloud, 13 Ads, 2 Analytics. Take the Ads subset and you take 13 skills, not 111.

Why one number cannot work here

The documented figure over-predicts on all three, but by wildly different amounts, and the size of the miss does not order by description length. That is worth stating carefully because it is where I first got this post wrong.

Pack	Median description	Per skill	Documented over-predicts by
mattpocock/skills	149 chars	34.1	193%
addyosmani/agent-skills	249 chars	95.9	4%
google/skills	397 chars	81.3	23%

If cost simply tracked description length, the miss would shrink monotonically as descriptions got longer. It does not: the middle pack is the closest to the documented figure and the longest is not. google/skills has descriptions 1.6x longer than addyosmani/agent-skills and costs 15% less per skill.

This site has published that already, from these same three packs: per-skill cost does not order by description length. Length is clearly involved — the shortest-description pack is by far the cheapest and by far the worst fit to the documented figure — but it is not the whole mechanism, and I am not going to invent the rest from three points.

So the practical version is deliberately modest: ~100 tokens per skill held as a ceiling across every reliable pack here, whose median descriptions ran 149 to 397 characters. How far under ran from 4% to 193%, and nothing measured here predicts which you will get. Above 397 characters this site has no reliable data at all, so do not assume the documented figure holds there. I am deliberately not reaching for the single-skill packs to argue otherwise, even though two of them have descriptions past 397 characters and measured well over 100, because packs I have excluded from every other claim cannot be evidence when they happen to suit one.

And the part that goes against me. This site published its own model for the same quantity, roughly 13.7 tokens plus 0.276 per character. If you are going to recommend measuring over trusting a figure, the honest thing is to check your own figure too:

Pack	Measured	Documented ~100	This site’s rate	Closer
addyosmani/agent-skills	2,302	2,400 (4% out)	2,124 (8% out)	the docs
google/skills	9,020	11,100 (23% out)	14,699 (63% out)	the docs, indicative only
mattpocock/skills	852	2,500 (193% out)	1,397 (64% out)	this site

Anthropic’s rough one-liner beats this site’s fitted model on two of the three packs, though one of those two is indicative rather than clean: this site published its rate as valid for descriptions of 30 to 300 characters, and google/skills has a 397-character median, so there the rate is being run outside its own stated range. The rate only wins where the documented figure is at its worst. That is a good reason to be modest about the whole exercise, and the reason the recommendation is “install it and measure it” rather than “use my number instead”.

What I am not claiming

Not that the documentation is wrong. It says “~100” with a tilde, in a table whose purpose is to convey that Level 1 is cheap and Level 3 is free until read. For that purpose it is fine. What it cannot do is answer “what will this pack cost me”, which is the question someone installing 111 skills actually has.

Not that the one-skill packs disagree. Three of the six packs measured are single skills and came in at 117, 262 and 274, above the documented figure rather than below. Re-verified 2026-08-18: even that aside was more than the data supports. Twelve further readings of diagram-design on Claude Code 2.1.234 ran from 70 to 967 tokens, so the same pack sits below the documented ~100 as readily as above it. Saying they came in “above rather than below” drew a direction out of figures this post correctly calls non-evidence in the very next sentence. There is no direction to draw. See flag-effect-2-1-234.json. Those are not evidence and I am not using them. This site has already established that a single skill cannot be resolved against the startup floor: one round of the hallmark pack measured 404 tokens below an empty floor, which installing a skill cannot do. They are published in the corpus and excluded from every claim here.

Then I asked six model runs the same question

Since the figure is documented but sits in a table most people will not read, I wanted to know what someone gets by simply asking. One ordinary question to duck.ai, free tier, no account, a New Chat each time so no answer could contaminate the next: “How many tokens does each Agent Skill add to Claude Code’s startup context?”

Model, as the UI labelled it	Answer
GPT-5.4 nano	~100 tokens
GPT-5.4 mini	~100 tokens, with a citation to the vendor page
GPT-5.4 mini (2nd run)	~100 tokens
Claude Haiku 4.5	Declined
Mistral Small 4	50–200 tokens
Mistral Small 4 (2nd run)	Declined

The three runs that gave ~100 were right to, and the citation was genuine. I checked the cited page rather than assuming a fabrication, which is the mistake I was set up to make: this site has published that only 6 of 30 AI-generated citations were real, so a hallucination was my expectation. It was not one.

Claude Haiku 4.5 declining is the interesting one, because the answer is documented by the company that makes it. It suggested four routes, one of which was “Testing empirically by comparing the startup context token usage with and without specific Agent Skills enabled” — which is how the numbers at the top of this post were produced.

And the same model, asked the same question in a fresh chat, gave the opposite kind of answer. Mistral Small 4 gave a 50–200 range on one run and declined on the next. That undercuts the whole table: if one model can move between those two states on identical prompts, a single answer from any of them is not evidence of anything.

An earlier draft of this post listed four specific errors in that Mistral answer. Two did not survive checking and I withdrew them; then I found I could not properly evidence the remaining two either, so they are gone as well. One rested on truncating its quote — it said the exact count is undocumented, and what Anthropic publishes is ~100, explicitly approximate, so on the full quote it is defensibly right. Another compared its figure for a base system prompt against this site’s whole-session startup floor, which includes tool definitions: different quantities. Of the two that remained, I had not preserved the verbatim text needed for a reader to check one of them. A public accusation against a named company’s model that a reader cannot verify from the published record should not be made, so none is made here.

Treat that table as an impression, not a measurement, and it is labelled that way in the corpus. Six runs across four models cannot characterise a model, and this trial demonstrated its own limitation by contradicting itself. The numbers at the top of this post are the measured part.

What this does not settle

Six packs, one day, one machine. Claude Code 2.1.233, macOS. The packs are pinned by commit in the source corpus.

Reliable for multi-skill packs only. Three of the six are single skills and are excluded from every claim, for the reason given above.

The model trial is n=6 and unstable, as above. It supports no claim about any model.

Description length is involved, but it does not order the result. google/skills has descriptions 1.6x longer than addyosmani/agent-skills and costs 15% less per skill. Something else is also moving, and this site has not established what.

Controls

The measured figures are not new here. They come from skillpack-cost-2-1-233.json, 56 runs across two harnesses, which has already been through two adversarial verification passes and one published correction. This post computes the doc comparison from that file rather than retyping the numbers, so the two cannot drift apart.

The vendor figure was read, not remembered. Quoted verbatim from the page on 2026-08-15 and linked. Worth noting the Claude Code skills page carries no per-skill token figure at all; its only token numbers are auto-compaction budgets. The ~100 is on the platform docs.

Model labels come from the UI, screenshotted. Recorded from the label printed above each answer rather than scraped from the DOM, because a product can change what sits behind a label without saying so.

One New Chat per model run, so no previous answer sat in context.

Nothing scripted. Six prompts typed by hand into a public UI on its free tier.

Every model run, with its verdict and the errors in it, is published in skill-token-doc-vs-measured.json.