Does Claude Watermark Its Text?

If you paste text from Sonnet 5, Opus 5, or Opus 4.8 into a checker today, an honest tool will tell you the same thing: Unknown. There is no public detector for Claude's watermark yet, and independent testing on August 23, 2026 found no watermark in current Claude output.

Here is what Anthropic committed to, in its own words, and what that means for anyone using Claude for writing, code, or client work.

What Anthropic announced

On August 10 and 11, 2026, Anthropic published a help center article titled "How Claude marks AI-generated content." Three days later, the company posted a companion blog post explaining the mechanism. The two documents are the primary source for everything below.

Anthropic's commitments:

1. Claude models launched on or after August 2, 2026 support machine-readable marking at launch.
2. Existing models are "in progress" for retrofit.
3. Marking applies across Claude Platform (API), Claude, Claude Code, Cowork, and Claude Tag, including Claude accessed through AWS, Google Cloud, and Microsoft Foundry.
4. Two mechanisms: an embedded watermark in generated text, and signed C2PA provenance metadata attached to generated files (SVG, PNG, JPG).
5. A watermark detection API is coming. No date was given.

The trigger for all of this is the EU AI Act's Article 50, which became enforceable on August 2, 2026, and requires providers of generative AI to output content that is "marked in a machine-readable format and detectable as artificially generated or manipulated." Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content, and applies the marking globally rather than only in the EU.

What is marked, and what is not

The watermark applies to generated text. Anthropic states clearly what it is and is not.

Marked: original text output from Claude, translations by Claude, and comments inside generated code where wording is a choice rather than a requirement.

Not marked: places where Claude has no choice in wording. If the exact string is required, such as a function name, a factual constant, or a piece of code that would break if a different token were picked, the watermark is skipped for that token.

Weakened, but still present: text that is lightly edited or paraphrased. Anthropic says light editing "probably won't remove the watermark completely; a complete rewrite where every word is replaced will."

One line from Anthropic's blog is the sentence that undoes most of the "Claude watermark remover" claims on the web today:

"Nothing is added to the text and there are no hidden characters."

That sentence appears in the design principles section of the August 14 post. It means the mark is not a zero-width space, not a narrow no-break space, and not any other invisible Unicode character. Any tool that offers to strip Claude's watermark by scanning for hidden characters is scanning for something else. It may still be a useful cleanup tool, since some models and pipelines really do insert invisible characters and any pasted text can carry them from web clipboards, but what it removes is not the Claude watermark.

Which Claude models actually carry it today

This is where most of the coverage got ahead of the facts.

Anthropic's rule is model launch date. Every Claude model shipped as of August 30, 2026 predates the August 2 cutoff:

• Sonnet 4.6 (February 17, 2026)
• Opus 4.7 (April 16, 2026)
• Opus 4.8 (May 28, 2026)
• Fable 5 and Mythos 5 (June 9, 2026)
• Sonnet 5 (June 30, 2026)
• Opus 5 (July 24, 2026)

No Claude model launched on or after August 2, 2026 has been announced or shipped as of this writing. Retrofits for existing models are "in progress" with a legal deadline of December 2, 2026 for systems that were on the EU market before Article 50 became enforceable.

An independent test on August 23, 2026 fed prompts to Sonnet 5 and Opus 4.8 through the API and found no statistical watermark signal in the output. The same test found SynthID present in Gemini 3.5 Flash. That is one data point, and Anthropic has not published the exact scheme needed for anyone outside the company to verify. But it is consistent with what Anthropic committed to: models launched after the cutoff carry it, and older models will carry it once retrofitted.

For the current state, the status tracker shows a verdict per provider with the date it was last checked, and the per-model list does the same for Claude specifically.

What Claude will say about your text once the detector exists

Anthropic has been careful about what a watermark verdict means:

"A watermark only helps test whether Claude might have produced or processed the content. It doesn't say anything about ownership or authorship and doesn't change a user's rights."

The mark shows that Claude processed the text. That includes proofreading, translating, summarizing, or converting a document. It does not identify a user, a chat, or an organization. Anthropic has been explicit that watermarking "carries no identifying information and can't be traced to a specific person, organization, or chat."

So the verdict from a future detector will be closer to "Claude was involved in producing or processing this text" than "this student wrote this with AI." That distinction matters for teachers, editors, and anyone worried about false positives.

What to do right now

If your question is whether Claude left a watermark on the essay you just generated, the honest answer for August 30, 2026 depends on which model.

For Sonnet 4.6 through Opus 5: probably not, based on Anthropic's cutoff rule and the one public test that has been run. Verdict: Unknown, leaning Clean.

For any new Claude model shipped since August 2, 2026: it should be marked at launch. Verdict: Unknown until the detection API opens.

For anyone in the EU using a Claude model that predates the cutoff: Anthropic must retrofit it by December 2, 2026. Assume the mark will land quietly.

You cannot check the mark yourself yet. Nobody outside Anthropic can. When the detection API ships, it gets wired up here the day it opens and noted on the status page. Until then, any site claiming to detect the Claude watermark is guessing or checking something else.

The one honest thing you can do today is remove things that really are removable: hidden Unicode, clipboard HTML fragments, and metadata on generated images. The Claude watermark detector runs those checks in your browser and labels the statistical row Unknown, with the reason attached.