A November 3 entry in the Claude Platform release notes adds three models to the list: Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carry Anthropic’s text watermark. I wrote about how the watermark works back in August.
The catch sits in a footnote
Sonnet 4.6 comes with a caveat, and it is a practical one. If a request sets
temperature, top_p or top_k to a non-default value, the output might not carry the
watermark at all. Anthropic’s advice: if you rely on the watermark, omit all three
parameters, “including any that a framework or client library sets for you”.
That is the awkward part. Plenty of SDK wrappers and agent frameworks set temperature
on your behalf without ever mentioning it.
Who is covered now, and who is not
The help centre keeps a table. Text watermarks on Anthropic’s own surfaces and through the cloud partners now cover Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5, 5, 4.8 and 4.7, Sonnet 5.5, 5 and 4.6, plus Haiku 5.5.
Still without a text watermark: Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. Those get Content Credentials under the C2PA standard on generated files and nothing else. The rollout on Amazon Bedrock is due to finish by November 10.
Why Anthropic is doing this
The driver is the EU AI Act’s Article 50(2) Code of Practice on transparency, which Anthropic has signed. Models launched in the EU on or after August 2, 2026 support marking from day one. Older models get a transition period, and that is the backlog being cleared right now.
Technically the model weaves the mark into the text itself. It adds no tokens and no hidden characters, changes neither meaning nor readability, and carries nothing about you or your organisation. Because it is part of the text, it survives copy and paste and some editing.
Detection is still a private preview. Access goes to regulators, law enforcement, media, fact-checkers, researchers, educational organisations and EU civil society groups, plus enterprises that have to verify marks on their own Claude output for compliance reasons.
If you depend on it, audit your sampling settings
For most people this entry changes nothing. Two groups should look closer. If you ship Claude-generated text in a product and have to demonstrate to an auditor that it is marked, walk through the sampling parameters in your stack once. And if you were hoping to use the watermark to spot AI text, the documentation hands you the sober answer: a detected mark is a signal, not proof, and a missing one tells you even less.