The Claude text watermark now reaches older models. Anthropic’s release notes say text from Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries its invisible watermark, in an entry dated November 3, 2026.
Anthropic started marking text with Fable 5.1 and Mythos 5.1 on September 1, so this entry extends the mark to earlier models. If you draft with Claude through the API, a cloud partner or the apps, the paragraphs you paste elsewhere can now be checked against Anthropic’s key. The company’s help article describes no way to turn that off.
TL;DR: Anthropic’s November 3 release note adds Mythos 5, Opus 4.7 and Sonnet 4.6 to the models that watermark their text. The mark changes word choices, not visible characters, and it can survive light editing. Anthropic’s help page lists no opt-out, and detection is limited to a private preview, so the mark is a probability signal, not proof of authorship.
Which models carry the Claude text watermark
Anthropic’s release notes list two rounds so far. Fable 5.1 and Mythos 5.1 got the mark on September 1, and Mythos 5, Opus 4.7 and Sonnet 4.6 followed on November 3.
The help article also lists newer models such as Opus 5.5, Sonnet 5.5 and Haiku 5.5 as marked, and says marking applies on Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag, wherever Claude is offered worldwide. The rollout on Amazon Bedrock for Mythos 5, Opus 4.7 and Sonnet 4.6 is due to finish by November 10. If you track which model is current, our AI model tracker keeps the lineup in one place.
| Model | Text watermark | Notes |
|---|---|---|
| Fable 5.1, Mythos 5.1 | Yes | Added September 1, 2026 |
| Mythos 5, Opus 4.7, Sonnet 4.6 | Yes | Added November 3, 2026 |
| Fable 5, Opus 5, Opus 5.5, Opus 4.8 | Yes | Listed in the help article |
| Sonnet 5, Sonnet 5.5, Haiku 5.5 | Yes | Listed in the help article |
| Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 | Not listed | No text watermark in the help article |
That last row is where the “older models” label stops being a clean description.
How the watermark works without changing your text
Anthropic’s explanation of the watermark, published August 14, says Claude normally picks each word from several plausible options using a random number. The watermark swaps that random number for a secret key plus the preceding words.
Anyone holding the key can then test whether a passage fits the pattern, which yields a probability that Claude wrote it. Nothing visible is added, and Anthropic says the mark does not affect quality or add tokens, so cost and speed stay essentially the same.
For a writer, nothing on screen changes. The pattern lives in word choice, so it travels with copied text.
Where the mark can fade or fail
Anthropic says light editing probably will not remove the watermark, while a complete rewrite where every word is replaced will. Short passages are unreliable, and factual text with few free word choices carries a weaker signal.
Sonnet 4.6 has its own catch. Anthropic’s developer documentation says a request that sets temperature, top_p or top_k to a non-default value might not carry the mark, and advises omitting all three if you rely on it, including values a framework sets for you.
The help article adds that a missing mark does not prove text was written by a person. For most people this will be invisible, until the one time someone treats a missing or present mark as a verdict.
Detection is narrow and there is no off switch
Text detection is in private preview for eligible groups such as regulators, researchers, media, fact checkers and enterprises checking their own output. The help article says Anthropic plans to widen access over time.
Anthropic has signed the EU AI Act’s Article 50(2) transparency code of practice, and the help article tells developers to assess for themselves what Article 50 requires of their products. The help article describes no opt-out for the watermark, and the developer page does not mention one either.
What to check if your work runs on Claude text
If you publish text drafted with Claude, assume it may be detectable by an approved checker, and that a clean result will not settle who wrote what. Anthropic’s own wording is that the mark cannot tell “Claude wrote this” from “Claude heavily edited this.”
Developers on Sonnet 4.6 should look at their sampling parameters, and anyone on an older model should check the model table before assuming either way. Our report on SynthID in iOS 27.2 shows a similar idea on the image side.
