Four older Claude models get a text watermark, and two of them skip it if a sampling setting changes
The Decoder, on Saturday, says an email to API customers names Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. On the last two, the mark applies only at the default temperature, top_p and top_k. FourWeekMBA, on Sunday, says a screenshot of the email gives 3 November as the start. Anthropic’s August post explains the mark as a pattern in word choice, not hidden characters.
Feature4 minEasy
Maya Chen4 min read
- Anthropic
- Claude
- Watermark
- EU AI Act

Photo: The DecoderSource page
Anthropic has told API customers it will put its text watermark on Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Matthias Bastian at The Decoder, in a story dated 31 October, says the notice was an email, that newer models such as Claude Opus 5.5 already carry the mark, and that the rollout on Amazon Bedrock may take a few extra days. Gennaro Cuofano at FourWeekMBA, writing Sunday, says a customer email shared on X that day is headed “Text watermarking is coming to more Claude models,” and that it states: “On November 3, 2026, Anthropic will begin applying its text watermark to outputs from these four models.” Cuofano writes that he has seen the email only as a screenshot, and that Anthropic has not published it. Bastian’s account does not print a date.
The Decoder’s account includes a limit Cuofano’s does not. On Opus 4.6 and Sonnet 4.6, Bastian writes, the watermark applies only when temperature, top_p and top_k stay at their defaults. Those three settings control how much randomness the model uses when it picks the next word. Changing any one of them, he writes, is enough to avoid the mark. The email, as he describes it, does not attach that condition to Mythos 5 or Opus 4.7. He reads the omission as a suggestion that the mark applies to those two whatever the sampling settings are. The lines he prints do not say that outright.
If you rely on the watermark for your own compliance, keep these settings at their defaults.
— Anthropic, in an email to API customers, as quoted by The Decoder
A pattern in the choice of words, not a hidden string
Anthropic explained the method on 14 August, in “How Claude’s text watermark works,” and updated the detection section on 1 September. A model writes one word at a time and chooses among candidates. After “The weather today was cold and…”, the company says, “overcast” and “grey” are both plausible and “sugary” is not. Where two words would do, a random number usually settles it. The watermark keeps the choice random and changes where the randomness comes from: a key, plus a few words that came before, settles which word is picked. Someone with the key can later ask how likely it is that the sequence matches choices Claude would have made. Anthropic says nothing is added to the text, there are no hidden characters, the mark does not require extra tokens, and it will not cost more.
Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat.
— Anthropic, “How Claude’s text watermark works,” 14 August 2026
The August article says the method is a version of SynthID-Text, which Google DeepMind published in Nature in 2024, in a line of approaches that goes back to a proposal by Scott Aaronson in 2022. Anthropic says its own tests found no impact on content, creativity or readability. It cites a DeepMind test that served a watermarked model to part of Gemini’s traffic and found no statistically significant difference in thumbs-up and thumbs-down ratings, and a study in which people comparing answers side by side saw no difference in quality. Those are tests the company describes. They are not a new measurement of Opus 4.6 or Sonnet 4.6. Bastian notes two objections he did not run himself: the blogger John Gruber’s argument that no two synonyms mean the same thing, so a key can cost precision, and a concern from the legal outlet Artificial Lawyer about clients who have banned AI use, and about more than one watermark in one document.
Where there is only one right word
The mark is thinner where the next word is not a free choice. After “Isaac Newton’s most famous work was called Principia…”, Anthropic writes, the next word has to be “Mathematica,” so the watermark has nothing to act on. The same limit applies to a proofread that changes only grammar and punctuation, and to code that has to be exact. A comment inside code can carry the mark, the article says, because a comment can be phrased more than one way. A translation carries it, because every word is Claude’s. Light editing probably will not remove the mark completely. A rewrite that replaces every word will. Detection answers only how likely it is that Claude was partly involved. It does not show that a person did not write the text, and it does not name a different model.
Who can check is a shorter list than who receives the mark. The detection API is in private preview for organisations Anthropic says EU law requires it to serve, including regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil-society groups, and for enterprises that have to verify the mark for their own compliance. The company says it applies the mark globally because it does not yet have a durable way to limit it by region. The reason it gives is the EU AI Act. It says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, with around 190 signatories, and that the code requires a method of marking generated text. For models launched before 2 August 2026, the August article says the law includes a transition period. The customer email is that later rollout. Bastian writes that customers cannot turn the mark off, and that, because it sits in the model, it works on Amazon Web Services, Google Cloud and Microsoft Foundry. Cuofano writes that a Help Center table he counted this week does not list a cloud-partner mark for these four models, while the email he saw says Bedrock may take a few more days. The email Anthropic has not published is the document that would settle the date, the sampling exception and the cloud timing.


