Claude watermark comes to Mythos 5, Opus 4.7 and Sonnet 4.6

Anthropic has extended the Claude watermark to older models: as of November 3, text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic’s invisible watermark. The news landed in the Claude Release Notes, together with one caveat about Sonnet 4.6.

What exactly changed?

A text watermark isn’t new in itself. According to the Release Notes, Claude Fable 5.1 and Mythos 5.1 have had it since September 1, and every model released in the EU since August 2, 2026 gets it on launch day. Older models were given a transition period, and Anthropic is adding them gradually.

Now three more models join the group. According to the help page on content marking, the watermark works on all Anthropic surfaces (Claude Platform, the Claude apps, Claude Code, Claude Cowork and Claude Tag) and with cloud partners. For Mythos 5, Opus 4.7 and Sonnet 4.6, the rollout on Amazon Bedrock is due to finish by November 10.

The same page lists full support for, among others, Fable 5.1, Mythos 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.

How does the Claude watermark work?

The mechanism is clever. Anthropic explains it in a separate post: the model generates text token by token, and when several candidates are roughly equally good, a random number usually breaks the tie. The watermark replaces that number with a different one, computed from a secret key and the preceding words.

To a reader, the text looks completely normal. But someone who has the key can check whether a sequence of words fits the pattern and estimate the probability that Claude wrote it. The method is a version of SynthID-Text, which Google DeepMind described in Nature in 2024, and the whole family of such methods goes back to a proposal by Scott Aaronson from 2022.

According to the documentation, the watermark:

1. doesn’t change the meaning, quality or readability of responses,
2. adds no tokens and no hidden characters, so the cost stays the same,
3. contains no information about us or our organization.

We don’t need to change anything in our requests or in how we handle responses.

What should we watch out for?

First, sampling parameters. On Sonnet 4.6, text generated with a non-default temperature, top_p or top_k may not carry the watermark. If we rely on it, the documentation advises omitting all three parameters, including those a framework or library sets for us. The help page lists the same limitation for Opus 4.6.

Second, the length and type of text. According to Anthropic, detection works poorly on short passages, and factual text and code, where there is usually one correct word, carry less of the mark. Light editing is unlikely to remove it, but a full rewrite will. Interestingly, translations done by Claude carry it too, because the model chooses every word in them.

Third, what the watermark means. A detected mark is a signal that Claude processed the text, not proof that it wrote it: it may only have corrected, summarized or translated it. The absence of a mark doesn’t settle anything either.

Fourth, access to detection. The detection API is in private preview, for regulators, media, fact-checkers, researchers and educational organizations, among others. The help page doesn’t mention any way to turn the watermark off.

Where did this come from?

From the EU AI Act. Anthropic signed the Code of Practice for Article 50(2) on transparency of AI-generated content, and since August 2, 2026 the EU has required AI providers to mark such content. Files (e.g. PNG and JPEG) are marked differently by Claude, with signed metadata in the C2PA standard. I wrote about how Brussels started enforcing the rules against the big labs in my article on AI Act enforcement.

It’s another change to the Claude API in recent days, after message threads.

FAQ

Does the Claude watermark make responses worse?

According to Anthropic, no. In internal tests the company measured no impact on content, creativity or readability, and the watermark adds no tokens.

How can we check whether Claude wrote a text?

For now, we can’t do it ourselves. The watermark detection API is in private preview for selected organizations, and Anthropic plans to expand access gradually.

Which Claude models have the watermark?

According to Anthropic’s help page, among others Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5, and since November 3 also Mythos 5, Opus 4.7 and Sonnet 4.6.

Can the Claude watermark be removed?

According to Anthropic, light editing is unlikely to remove it, but a complete rewrite of the text, word by word, will.