Claude Text Watermark: Power-User Guide to Models, Sampling and Detection

Last updated: November 17, 2026

Claude’s text watermark is an invisible statistical pattern that Anthropic weaves into text from supported Claude models. On November 16, 2026, Anthropic extended it to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. It adds no tokens or cost. API users of those models who need the mark should leave temperature, top_p and top_k unset.

Contents

What changed on November 16

The Claude Platform release notes for November 16, 2026 say text generated by Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5 now carries Anthropic’s text watermark on supported platforms. If a request to one of these models sets temperature, top_p or top_k to a non-default value, the text might not carry the watermark.

This is the next step in a rollout Anthropic described in its help center article on marking AI-generated content. Claude models launched in the EU on or after August 2, 2026 support marking at launch. For older models, the EU AI Act allows a transition period, and Anthropic is adding marks to them over time.

Which models and platforms are covered

The table below summarizes Anthropic’s published support list as of November 17, 2026.

Model group Text watermark (Claude apps and API) Cloud partners (AWS, Google Cloud, Microsoft Foundry) Sampling caveat
Claude Fable 5.2, 5.1, 5; Mythos 5.2, 5.1 Yes Yes No
Claude Mythos 5, Opus 4.7 Yes Yes; Amazon Bedrock rollout may take a few more days No
Claude Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5, Haiku 5.5 Yes Yes No
Claude Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5 Yes Yes; Bedrock rollout may lag Yes
Claude Opus 3 (access on request) Yes First-party surfaces only Yes

Anthropic says marks apply wherever supported models run, including the Claude Platform API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, in every region. Because the watermark is applied at the model level, it is present no matter which product the text comes from. The help center table also marks every listed model as supporting Content Credentials (C2PA) on files.

For the newest model in that list, see our Claude Fable 5.2 guide.

Sampling settings that can drop the watermark

The Text watermark page in the Claude Platform docs says you do not need to change your requests or response handling. The one exception applies to Claude Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 and Opus 3. If you need output from those models to carry the watermark, follow these steps:

  1. Find every place your code calls one of those six models through the Messages API.
  2. Remove temperature, top_p and top_k from the request body. Anthropic says a non-default value for any of them may leave the output unmarked.
  3. Check your framework or client library. The docs warn that some set these parameters for you, so inspect the request body it actually sends.
  4. Re-run a sample request and confirm the logged body no longer contains the three parameters.
  5. If you have compliance duties, record which model and settings each workload uses, since Anthropic lists the sampling caveat only for those six models.

Some teams set a low temperature for repeatable output. For those older models, that choice now trades off against reliable watermarking, so decide which matters more for each workload.

How the watermark works

Anthropic’s explainer, How Claude’s text watermark works, says the method is a version of the SynthID-Text approach that Google DeepMind published in Nature in 2024. When the model chooses between equally good next words, a secret key and the preceding words settle the choice instead of an ordinary random number. Anyone holding the key can then estimate how likely it is that Claude wrote the text. Our explainer on SynthID covers the original technique.

  • No cost or token change. No extra tokens are produced, so the price is the same, and the speed impact is negligible.
  • No personal data. The watermark cannot be traced to a user, organization or chat.
  • Global. Anthropic applies it worldwide because it does not yet have a durable way to scope it by region.
  • Why now. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, one of about 190 signatories.

How to check for Claude’s marks

Text and files are checked differently:

  • Text. The watermark detection API is in private preview. It is open to eligible organizations, such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus select enterprises that must verify marks on their own outputs. Others can register interest through the Claude Watermark Detector Access Request Form.
  • Files. Supported files such as PNG and JPEG get a signed C2PA Content Credential. Anthropic offers a free Claude Content Checker for those, and any C2PA-aware tool can read them.

Watermark, Content Credentials and AI detectors

Question Text watermark Content Credentials (C2PA) Third-party AI detectors
Where it lives In the word choices Signed metadata in the file Nothing is added; the tool guesses from phrasing
Who can check Holders of Anthropic’s key (private preview API) Anyone with a C2PA tool or the Claude Content Checker Anyone with the tool
Main weakness Short, factual or heavily rewritten text Stripped by re-saving, conversion or screenshots No access to Anthropic’s key

Other providers are adding their own marks under the same EU code. See our Grok API watermarking guide and whether ChatGPT will watermark its text in the EU.

Limitations

  • A match is not proof of authorship. It shows Claude was likely involved, which includes editing, translating or summarizing someone else’s work.
  • No match is not proof of human writing. Very short passages, heavy paraphrasing, translation by another tool, unsupported models and the sampling caveat above can all leave no detectable mark.
  • Editing. Light editing probably will not remove it; a complete rewrite will.
  • Proofreading and facts. When Claude only fixes grammar, or a sentence has one correct next word, there is little for the watermark to act on. Translations Claude produces are watermarked.
  • Your obligations. Anthropic says businesses deploying Claude should assess for themselves what Article 50 of the EU AI Act requires of them.

Frequently asked questions

Does Claude watermark its text?

Yes. Text from supported Claude models carries an invisible Anthropic watermark. Models launched on or after August 2, 2026 have it at launch, and on November 16, 2026 Anthropic added it to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5.

Can I see the Claude watermark?

No. Anthropic says the watermark adds no tokens or hidden characters and does not change the meaning, quality or readability of the text. It is a statistical pattern in word choices that only a detector with Anthropic’s key can check.

How do I check if text was written by Claude?

Anthropic’s watermark detection API is in private preview for eligible organizations such as regulators, media, fact-checkers, researchers and educators, plus select enterprises with EU AI Act obligations. Others can register interest through Anthropic’s access request form.

Does Claude watermark code?

Only lightly. Anthropic says code usually has to be exact, so there are few free word choices for the watermark to use. It can appear in places with arbitrary choices, such as code comments, with a negligible effect on the code itself.

Does the watermark cost extra or slow Claude down?

No. Anthropic says watermarking produces no extra tokens, so the price is the same, and it has a negligible effect on speed. It also carries no information about you, your organization or your chats.

Sources

Image: Anthropic logo image from Anthropic’s announcement page, used for editorial purposes.

Related guide: Claude Linked billing guide.

Related guide: ElevenLabs Content Provenance API guide.

Related answers

1 thought on “Claude Text Watermark: Power-User Guide to Models, Sampling and Detection”

Leave a Comment