Anthropic adds text watermarks to four Claude models

According to a Juejin report published on November 1, 2026, Anthropic has announced the addition of text watermarks to four Claude models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Anthropic describes the watermark as an imperceptible pattern in word choice that does not affect semantics, quality, or readability, and says it carries no user identity information — in the company's wording, it contains nothing about the user, their organization, or their conversations with Claude. The mechanism marks the fact that a text was generated by AI; it is not designed to trace the specific generator.

The report notes that the watermarking on the two older models, Claude Opus 4.6 and Claude Sonnet 4.6, carries a precondition: it only takes effect when developers keep three sampling parameters at their default values. Any developer who adjusts those parameters — for whatever purpose — may end up with unwatermarked output. Anthropic reportedly reminds customers in its documentation that anyone whose compliance work depends on the watermark must leave the sampling parameters at their defaults. The report does not name the three parameters or state their default values, leaving developers to check the official documentation. Whether the newer Claude Mythos 5 and Claude Opus 4.7 carry the same restriction is not mentioned in the source. The report adds that Opus 4.6 still ranks as Anthropic's best writing model in the company's internal evaluations, above Opus 5.5, but offers no explanation of any connection between that standing and the watermarking caveat.

The report also says Anthropic ties this retroactive watermarking of older models to a December deadline under the EU AI Act, though that section of the original text cuts off mid-sentence. This remains the sole source on the story, and further reporting is needed to establish what followed. The article is silent on when exactly Anthropic published the watermarking documentation, what tools or API endpoints expose it, and whether image, audio, or video outputs from these models are covered by the same mechanism.