Mashable reported that Anthropic had scheduled text watermarking for Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 to start on November 3, 2026. In Anthropic’s support matrix inspected November 4, text-watermark support was not listed for those versions, while C2PA file credentials were. Anthropic says it plans to cover older models by December 2, 2026.
Four Claude versions were named in the reported schedule
The reported schedule covered Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. The November 3 date was reported as a planned start, rather than evidence that marking had begun on each version.
Text watermarks and file credentials are different
Anthropic’s support matrix, inspected November 4, listed C2PA credentials for files from the four versions but did not list text-watermark support for them. Anthropic says all covered older models will have marking support by December 2, 2026. The company describes its marking commitments in connection with the EU AI Act’s Code of Practice on transparency of AI-generated content.
A text watermark is a pattern in a model’s word choices. Anthropic says its approach is based on a version of SynthID-Text. C2PA Content Credentials, by contrast, are signed metadata attached to supported files, including PNG and JPEG. One kind of mark does not indicate that the other is present.
A detected mark points to involvement, not authorship
Anthropic says a positive text-watermark result suggests Claude may have processed the text. It cannot establish that Claude wrote the whole passage or identify a person, organization or chat. Proofreading or editing a text with Claude, for example, may leave a mark even when the underlying writing came from someone else. Anthropic’s explanation of Claude’s text watermark describes that limit.
Text detection remains in private preview for eligible organizations and enterprises with relevant compliance obligations. Very short samples may not contain enough signal for reliable detection, and Anthropic has published no numeric minimum length.
The same constraints matter for code. Anthropic says code generally carries less watermarking because exact syntax limits word choices; comments and other flexible text may carry a signal.
