Claude’s invisible text watermark now covers three older models

Anthropic says text from Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries its text watermark, according to its November 3 Claude Platform release notes. On Sonnet 4.6, one set of API settings can strip it.

Three older models join the watermarked list

The release note is short. Output from the three models is marked on the platforms listed in Anthropic’s help article, How Claude marks AI-generated content. The newer Claude Fable 5.1 and Claude Mythos 5.1 were already covered: Anthropic’s September 1 release note said text from both carries the watermark, and that files made through its code execution tool get C2PA Content Credentials.

Anthropic’s developer documentation says the watermark does not change meaning, quality or readability. It adds no tokens or hidden characters and carries no information about the user or their organization. Developers do not need to change requests or response handling.

There is one exception. On Claude Sonnet 4.6, a request that moves temperature, top_p or top_k away from its default value may return text with no mark at all. The docs tell anyone who relies on the watermark to omit all three parameters, including any that a framework or client library sets automatically. Anthropic does not say how often the mark goes missing in that case.

What the watermark can and cannot prove

The mark sits in the wording itself rather than in file metadata, which is what separates it from C2PA credentials on images. According to Anthropic’s help article, it stays with the text when copied, but it may not survive all editing.

That article also sets out where the marks apply. Text watermarks cover Claude Platform (the API), Claude, Claude Code, Claude Cowork and Claude Tag, and also apply when the models are reached through AWS, Google Cloud and Microsoft Foundry. C2PA credentials apply only where file generation exists, in the Claude apps and the API.

Ordinary users cannot check text yet. According to the same article, watermark detection is in private preview for eligible organizations, and Anthropic plans to expand access. Anyone can check files with a Claude Content Checker, but that covers the image credentials, not the text mark.

A mark that suggests, not a verdict

The angle that matters here is how far a reader can trust the signal. Anthropic is careful on this point. A detected mark suggests Claude processed the content but is not conclusive, because Claude may only have edited or translated text someone else wrote, and the content may have changed afterward.

A missing mark proves even less. The help article lists older models, heavy editing, very short text, stripped metadata and unsupported platforms as reasons a mark can be absent. A school or newsroom that sees no watermark therefore cannot conclude that a human wrote the text, and the Sonnet 4.6 sampling caveat adds one more way for a mark to vanish unnoticed.

The timing also lines up with regulation. Per the help article, models launched in the EU on or after August 2, 2026 mark content from launch, and Anthropic is adding marking to earlier models. Mythos 5, Opus 4.7 and Sonnet 4.6 fall into that earlier group. Anthropic does not tie the change to any specific law in the release note, so the link to EU rules is context rather than a stated motive.

Anthropic has also been working on a separate route to provenance. Aiholics reported on FlashProve, a system that proves which model wrote a reply, at 72 times the cost. A watermark costs nothing extra but is probabilistic; a proof is stronger but expensive.

Who is affected now

Nothing needs to change for most API users, since marking is on by default for the listed models and platforms. Developers on Sonnet 4.6 who pass sampling parameters should check whether they need the watermark, because the docs advise dropping those settings. Developers who built tooling that assumes unmarked output from these models will now see marked text.

The release note does not say which cloud platforms or Claude apps carry the mark for each model, and refers readers to the help article for the current list. It also gives no detection accuracy figures, no minimum text length, and no timeline for opening watermark detection beyond the private preview. Those details come from Anthropic alone and have not been independently verified.