Anthropic expands text watermark coverage, with a Sonnet sampling limit
Anthropic said on November 3 that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries its text watermark on the platforms covered by its marking documentation. The release note expands the named model coverage; it does not establish that every Claude model or hosting platform is covered.
For people publishing AI-assisted work, the watermark is a provenance signal rather than a visible label. Anthropic says it does not change the meaning, quality or readability of the output, and adds no tokens or hidden characters.
Anthropic also says the watermark carries no information about the user or their organization. Developers do not need to change their requests or response handling for supported output to carry it.
There is an important exception for Claude Sonnet 4.6. Text generated by a request that sets temperature, top_p or top_k to a non-default value might not carry the watermark.
If an application relies on the watermark, Anthropic advises omitting all three sampling parameters. That includes parameters inserted by a framework or client library, so checking only the application's visible request settings may miss an override.
The announcement concerns marking generated text. It is not a claim that the text is factually correct, and it does not establish that a particular publisher has met every disclosure or legal obligation. Teams should check model and platform coverage before relying on the signal.