Mihata
AI Usage2026.08.18Updated 2026.11.23

Claude Watermark: Which Models Mark AI Text (Nov 2026)

Claude now hides an invisible, machine-readable watermark in the text it writes. If you use Claude for client work, school, hiring, or publishing, that changes what "AI-written" means in practice. This page sums up what Anthropic has actually published, as of November 23, 2026.

Short answer: Claude models released on or after August 2, 2026 carry a text watermark from launch, and on November 16, 2026 Anthropic extended it to older models including Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 (Claude Platform release notes). On those older models, setting temperature, top_p, or top_k to a non-default value means the watermark may not be applied. Files Claude generates, such as PNG or JPEG images, get signed C2PA Content Credentials instead.

What the watermark is

Anthropic describes two separate mechanisms (How Claude marks AI-generated content):

Output

How it is marked

What can remove it

Generated text

An imperceptible watermark woven into the text itself

Survives copy and paste. Anthropic says it "may persist through some editing", so heavy rewriting can weaken it

Generated files (e.g. PNG, JPEG)

Signed provenance metadata (C2PA Content Credentials)

Stripping metadata or re-capturing the image

According to Anthropic's developer docs, the text watermark does not change the meaning, quality, or readability of the output, adds no tokens or hidden characters, and carries no information about you or your organization (Text watermark). Because it lives in the wording itself rather than in a file, pasting Claude's answer into Word or an email does not strip it.

Which Claude models are watermarked (November 2026)

Models

Text watermark

Via AWS, Google Cloud, Microsoft Foundry

All other current models (Fable 5 family, Mythos 5 family, Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5)

Yes

Yes

Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5

Yes, with the sampling caveat

Yes, with the caveat (Amazon Bedrock rollout may take a few extra days)

Sonnet 4.6

Yes, with the caveat

Yes, with the caveat

Opus 3

Yes, with the caveat

Anthropic's own surfaces only

The list comes from Anthropic's help center table, which it says it will keep updating. Marking applies worldwide, not only in the EU, and across the Claude Platform (API), the Claude apps, Claude Code, Claude Cowork, and Claude Tag.

The sampling caveat for older models

For Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5, and Opus 3, Anthropic says that if a request sets temperature, top_p, or top_k to a non-default value, the output might not carry the watermark. If you need the watermark, omit all three, and check whether your framework or client library sets them for you (Text watermark). The practical flip side: an internal tool that runs an older model at a custom temperature may produce unmarked text.

Why Anthropic is doing this

Anthropic says it has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Article 50 requires providers of generative AI to mark synthetic output in a machine-readable, detectable way, with a transition period for systems launched before August 2, 2026. That date is why the model list is split where it is.

Can you check whether text came from Claude?

Not freely, for text. Watermark detection is in private preview for eligible organizations, such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups, plus select enterprises with obligations under the Act. Anthropic has an access request form and says it plans to widen access over time. For files, anyone can use the free Claude Content Checker to look for a Claude-issued Content Credential (Anthropic help center).

What it means at work

  • No watermark does not prove a human wrote it. Older models at custom temperatures, other AI vendors, and heavy rewriting can all leave text unmarked. Treat a missing mark as "unknown", not "human".
  • Disclose AI use by policy, not by hoping no one checks. Contracts, school submissions, and job applications that require disclosure are not affected by whether a detector exists today.
  • If your workflow must carry the mark (for example, to meet your own EU obligations), drop custom sampling parameters on older models, or move to a model released after August 2, 2026.
  • Nothing changes in your prompts or output handling. Anthropic says you don't need to modify requests or response processing for the watermark.

For how the newest Claude release compares with OpenAI's latest models, see GPT-6.2 vs Claude Fable 5.2.

Summary

Claude's text watermark is invisible, travels with copied text, and now covers every current Claude model, including Opus 4.6, Sonnet 4.5, and Haiku 4.5 since November 16, 2026. Older models can drop it when temperature, top_p, or top_k are changed. Text detection is limited to approved organizations; file credentials can be checked by anyone. A missing watermark is not evidence that a person wrote something.

FAQ

Does Claude watermark its text?

Yes. Claude models released on or after August 2, 2026 carry an invisible text watermark, and since November 16, 2026 it also covers older models such as Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5.

Does copying and pasting remove the Claude watermark?

No. The watermark is part of the text itself, so it travels with copied text. Anthropic says it may persist through some editing, which means heavy rewriting can weaken it.

Why might Claude output have no watermark?

On older models such as Opus 4.6, Sonnet 4.5, and Haiku 4.5, setting temperature, top_p, or top_k to a non-default value means the output might not carry the watermark.

Can anyone detect Claude-written text?

Not yet. Text watermark detection is in private preview for eligible organizations such as regulators, media, researchers, and educational institutions. Files can be checked by anyone with the free Claude Content Checker.

Feel free to contact us

Whether you have questions about AI, IT, or design, need a consultation,
or want to request a quote — don't hesitate to reach out.

Contact Us