Which Claude Models Are Watermarked Right Now

This page is reviewed weekly. Anthropic's watermark detection API is in private preview for eligible organisations as of September 8, 2026; when DetectMark gains access, each row moves from Unknown to Found or Clean, dated per check.

An independent test on August 23, 2026 found no watermark in Sonnet 5 or Opus 4.8 output.

The rule Anthropic committed to

From Anthropic's help center article (August 10 or 11, 2026):

"Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Existing models are in progress."

Two dates set the retrofit clock. August 2, 2026 is when EU AI Act Article 50 became enforceable. Anthropic signed the EU Code of Practice on Transparency and applied the marking policy globally, not only in the EU. December 2, 2026 is the EU deadline for pre-existing systems that were on the market before August 2 to reach compliance. Anthropic must have older Claude models retrofitted by then for EU users, and given the global policy, elsewhere in step.

Anthropic has not committed to a retrofit order between the older models.

Status table

The table lists notable current models, not an exhaustive registry. The stable fact is the rule above: models launched on or after August 2, 2026 support marking at launch, and earlier models are being retrofitted.

Every one of these models except Mythos 5.1 is reachable in Claude, the API, and Claude Code today. Mythos 5.1 is the same model as Fable 5.1 with fewer safeguards, limited to vetted organisations.

Verdict definitions used in this table and in every DetectMark scan verdict:

• Clean: the sample was long enough to score against a known detectable signal, and nothing was found.
• Found: a known signal is present in the sample.
• Unknown: the sample is too short, the model is not yet retrofitted, or no public detector exists for the check. Every Claude row is currently Unknown because Anthropic's detection API is in private preview, not public.
• Unsupported: the check does not apply.

The status matrix also labels provider-level capability, always with one of exactly four words: Available, Partial, Announced or Unavailable. Each is defined on the page.

Why every row reads Unknown

Unknown is the whole verdict. There is no public detector for Anthropic's mark, so nothing about these models can be verified from outside either way, and this page does not grade its guesses.

The evidence behind the verdict is still worth laying out. Fable 5.1 and Mythos 5.1 launched on September 1, 2026, after the cutoff, so under Anthropic's own rule they are marked at launch; Anthropic says so, and nobody outside its private preview can confirm it. Every earlier Claude model predates August 2, 2026, so none of those was marked at launch. That is not confirmation of anything today, because Anthropic could retrofit any of them at any time before the December 2 deadline.

One independent test on August 23, 2026 fed prompts to Sonnet 5 and Opus 4.8 through the API and looked for a SynthID-Text style statistical signal. It found nothing. The same test found SynthID in Gemini 3.5 Flash output. That test used the open-source SynthID-Text detector code, which cannot read Anthropic's key, so it is context rather than proof.

No similar test has been published for Sonnet 4.6, Opus 4.7, Fable 5, Mythos 5, Opus 5, Fable 5.1 or Mythos 5.1. Every row stays Unknown until DetectMark gains access to Anthropic's API or a credible third party publishes a reproducible check.

What has shipped since August 30

Post-cutoff models. Claude Fable 5.1 and Mythos 5.1 launched on September 1, 2026, the first Claude models released after August 2. Anthropic's announcement says the outputs of models released after that date carry the watermark.

The detection API, in private preview. As of September 8, 2026, Anthropic's help center describes watermark detection as available to eligible organisations (regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups, and enterprises with their own AI Act obligations) through an access request form. No public detector, and no date for one.

A free Claude Content Checker, which reads the Content Credential on files Claude generated. It checks the attached credential, not the text watermark.

The moment any of these changes, this page changes.

What triggers an update to this page

This page is reviewed weekly, and out of band whenever any of the following happens:

1. Anthropic launches a new Claude model. Row added, status set per the rule.
2. Anthropic announces a retrofit for a specific model. Row updated with the announcement date and, once verifiable, a run against the detection API.
3. Anthropic opens the watermark detection API. Every Unknown row becomes Found or Clean, run per model, dated.
4. An independent test on a specific model is published with reproducible methodology. Row updated with a note and a link.
5. The EU December 2, 2026 deadline arrives. Every EU-served model that has not been retrofitted becomes a compliance question, and the state gets noted here.

How to check your own text today

You cannot check for Anthropic's statistical watermark without Anthropic's key or the future detection API. Anyone offering that check today is guessing or checking something else.

What you can check today, in your browser:

• Hidden Unicode: the zero-width family, non-breaking spaces, byte-order marks, soft hyphens, joiners, variation selectors, bidi controls. Not the Claude watermark, but a real category worth cleaning.
• Lookalike characters: Cyrillic and Greek letters that render like Latin ones.
• Clipboard HTML fragments: class names and attributes that came along with a rich-text paste.

Paste text into the checker; the verdict is one of the four labels above; the clean copy is one click. Nothing leaves your device.

For images generated in Claude (SVG, PNG, JPG), the file carries C2PA signed provenance metadata. The image tool reads and reports it, and shows the assertions inside. That check is available today.

Legal notes

The EU AI Act Article 50 obligation sits with the provider, Anthropic, not with you. If your Claude model is not yet retrofitted, that is Anthropic's compliance question, not evidence of anything about your text.

Anthropic's own framing is worth keeping in view: "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." A future positive detection is not proof of dishonesty, cheating, or authorship. It is evidence that Claude touched the text.