Anthropic extends its text watermark to older Claude models, but the official sources don't match

The 3 November 2026 entry in the Claude Platform release notes announces that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on the platforms listed in the support page “How Claude marks AI-generated content”. The newest models already had the mark. What's new is the extension to some older models that are still on sale. The support page lists the covered surfaces as the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, plus the cloud partners AWS, Google Cloud and Microsoft Foundry. For Amazon Bedrock, however, Anthropic says the rollout to the three models will be completed by 10 November, so until then coverage on Bedrock should not be taken for granted. According to third-party reports, customers had been notified by email before 3 November. That email has not been verified directly.

Anthropic says it embeds “an imperceptible watermark directly into the text itself. You won't see it, and it doesn't change the meaning, quality, or readability”. The technical documentation adds that the watermark “adds no tokens or hidden characters and carries no information about you or your organization”. For generated files, where supported, the company uses C2PA Content Credentials instead. Anthropic does not say which technique it uses. The idea that it is a variant of SynthID-Text comes only from third-party sources. Nor has anyone verified how well the mark survives paraphrasing or translation.

The official sources don't fully agree with each other. The release notes name Mythos 5, Opus 4.7 and Sonnet 4.6, while the support page and third-party reports also include Opus 4.6. The same goes for sampling parameters. The release notes attribute only to Sonnet 4.6 the case in which a request with temperature, top_p or top_k set to a non-default value could produce unmarked text. The support page extends the same caveat to Opus 4.6. The model table on the support page also includes Opus 4.5, Sonnet 4.5 and Haiku 4.5, but it gives no dates and can't be read with certainty. A second wave on 16 November appears only in secondary digests and is unverified.

That leaves the question of who can read the mark. According to the support page, text detection is in private preview and reserved for eligible organizations: regulators, law enforcement, media, fact-checkers, researchers, educational institutions, EU civil society and companies checking their own outputs. The public Claude Content Checker deals with the Content Credentials of files, and the page doesn't say it checks text as well. The regulatory backdrop is Article 50 of the AI Act, which requires providers of generative AI to mark synthetic content in a machine-readable and detectable way. Anthropic says it has signed the Code of Practice on Article 50(2) and invites anyone building products on Claude to assess for themselves what the rule requires of their own services. Other providers are moving too: xAI has announced default watermarking on the Grok APIs from 7 November. The announcement on Anthropic's blog and the transitional December deadline tied to the Digital Omnibus are not verified.

For developers, the most concrete detail concerns parameters. The documentation tells anyone who wants the watermark to omit temperature, top_p and top_k, including those that frameworks and client libraries set on their own. A value inserted automatically by a library could therefore be enough to drop the mark on Sonnet 4.6 and, according to the support page, on Opus 4.6 as well. In the other direction, the support page mentions no way to switch the mark off, and Pasquale Pillitteri's independent article states explicitly that Anthropic doesn't allow it. It's worth checking your own calls before assuming your text is marked.

A mark nobody sees, and one that, for now, only organizations admitted to the preview can look for in text.

— Olya