Claude Code watermark: it's already in what it writes, and costs no tokens

Claude's watermark is already in the text Claude Code writes on current models, with no extra tokens and no price change. Where it lands (little in code, more in commits, PRs and docs) and what doesn't remove it.

TL;DR There's nothing to turn off and nothing to pay: the watermark adds no tokens or characters and doesn't identify you. If a PR or a README shouldn't come from Claude, write it yourself: neither removing Co-Authored-By nor running the text through another Claude rewrite takes the mark out.

Text written by the current Claude models carries an invisible watermark, and Claude Code is no exception. Anthropic applies it at the model level to meet the EU AI Act Code of Practice it signed in July 2026, worldwide and in every product: the API, the app, Claude Code, Cowork, and Claude models on AWS, Google Cloud and Microsoft Foundry. Its table lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 as marked already, the models Claude Code runs on today.

How it works

A model writes one word at a time, and often has two equally good options: after "the weather today was cold and…", both "grey" and "overcast" work. Without a watermark, a random number decides. With one, a key decides, a key only Anthropic holds, combined with the words that came before.

The text keeps its quality and meaning. Anthropic saw no difference in its testing, and its method is a version of SynthID-Text, Google DeepMind's, which showed none on live Gemini traffic either.
Nothing is added: there are no hidden characters to search for or strip.
With the key, you can work out how likely it is that Claude wrote a text. The longer the text, the more reliable the answer.

What it costs you

	With the watermark
Tokens	Not one more: it generates no extra tokens
Price	Unchanged
Speed	A negligible impact, per Anthropic
Your data	None: it can't be traced back to you, your organization or your chat
Turning it off	No setting: the Claude Code docs don't mention it, and it's applied in the model

Where it lands in what Claude Code writes

The watermark only lives in the words Claude picks when there's a choice to make, so it isn't spread evenly:

Code: little. When changing a token would break the program, there's no choice and the watermark isn't applied. Code carries less of it than other text.
Comments: yes. They're prose inside the code, with interchangeable words.
Commit messages and PR descriptions: yes. Free text Claude writes from scratch.
READMEs and docs: more, since they're long and detection gains confidence with length.
Translations: the whole text. Claude picks every word of a translation.
A light proofread of your own text: next to nothing. The words are still yours, and a handful of corrections may not be enough to detect.

What doesn't remove it

Removing Co-Authored-By: Claude. The attribution setting in settings.json strips that signature from your commits and PRs, a visible line the CLI adds. The watermark is in the words of the message, not in that line.
Running the text through another Claude rewrite. A skill like the humanizer changes the style patterns AI detectors look for, but Claude picks the new words too, so they're still marked. The same goes for translating an existing text with Claude.
Light editing. A few edits probably won't remove it completely. A full rewrite, word by word, will.

Files it writes to disk carry no C2PA

Anthropic also attaches Content Credentials (C2PA), a signed note in the metadata, to files Claude generates in its apps and in the API, such as a PNG or a JPEG. Claude Code doesn't: an SVG and a PNG written by v2.1.317 on Sonnet 5.5 came out without one, and the PNG holds nothing but its image chunks:

b'IHDR' 13
b'IDAT' 111
b'IEND' 0

The SVG has no metadata, neither file contains c2pa or jumbf, and the CLI binary has no C2PA code. To check a file generated in the app, Anthropic offers the Claude Content Checker.

Who can check for it

Nobody can run it on your repo at will today. Detection is a private preview API for organizations eligible under EU law (regulators, law enforcement, media, fact-checkers, researchers, educational organizations, civil society) and enterprises required to verify it. Tools like Pangram or GPTZero are something else: they look for the tells of AI writing and don't have the key.

And a detected mark only says Claude was involved, not that it wrote the ideas: it may have proofread, translated or summarized something of yours.

Official docs: How Claude marks AI-generated content · How Claude's text watermarking works

Requirements

Nothing to install or configure: the watermark is in the model. The file test ran on Claude Code v2.1.317 with Sonnet 5.5.