Anthropic is extending its invisible text watermark to four older Claude models starting today, November 3. Output from Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 now carries the mark, which Anthropic says lets a key holder estimate whether Claude was involved in producing a piece of text.
Anthropic told API customers about the change by email. Its newer models, including Claude Opus 5.5, already watermark their output, and models released after August 2, 2026 shipped with the feature at launch. The email said the rollout on Amazon Bedrock may take a few additional days.
The mark sits in word choice rather than in hidden characters. Anthropic describes it as an imperceptible pattern in the words a model picks when several are plausible, with a secret key and the preceding words steering those choices. The approach is a variant of Google DeepMind's SynthID text method.
Key facts
- 01From November 3, 2026, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 output carries Anthropic's invisible text watermark.
- 02On Opus 4.6 and Sonnet 4.6, the mark applies only when temperature, top_p and top_k stay at their defaults.
- 03Anthropic told API customers the Amazon Bedrock rollout may take a few additional days.
- 04The watermark adds no characters or tokens and does not change pricing or request and response formats.
- 05Text detection runs through a private-preview API for regulators, media, researchers and other eligible groups.
Anthropic says the watermark adds no characters or tokens, does not change pricing or the format of requests and responses, has little effect on output speed, and contains no information about users or their conversations. Existing integrations need no code changes.
There is one carve-out. On Opus 4.6 and Sonnet 4.6, the watermark applies only when the sampling parameters temperature, top_p and top_k are left at their defaults, so changing any one of them is enough to avoid it. The email told customers: "If you rely on the watermark for your own compliance, keep these settings at their defaults." It carried no such caveat for Mythos 5 or Opus 4.7.
The timing is tied to Article 50 of the EU AI Act, which requires marking of AI-generated content from August 2, 2026, with a December 2, 2026 deadline for systems already on the market. Anthropic applies the mark worldwide rather than only in the EU, and because it is embedded at the model level it follows Claude onto Amazon Web Services, Google Cloud and Microsoft Foundry. Customers cannot switch it off.
Detection is limited for now. Anthropic's text detection API is in private preview for organizations with verification duties under EU law, including regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus enterprises with similar obligations. Files such as PNG and JPEG images carry a signed C2PA Content Credential instead, which can be checked with the free Claude Content Checker at claude.com/check-content. That tool does not read text watermarks.
Anthropic is careful about what a result means. Its help center says "a detected mark provides a signal that content was processed by Claude, but is not fully conclusive," and that a missing mark does not prove content was human-written. Heavy editing, translation, short passages and stripped file metadata can all leave too little signal to detect.
OpenAI has taken a narrower route. Its textGrain system, launched in October 2026, is on by default only for ChatGPT and Codex users in the EU, while API customers worldwide can toggle it and it ships disabled.
The approach has drawn some pushback. Blogger John Gruber has argued that selecting words by watermark key rather than by meaning costs precision, and the legal publication Artificial Lawyer has warned that law firms whose clients ban AI use could have the mark surface in their work.
For teams on the affected models, the practical change is confined to the two 4.6 releases, where Anthropic's own guidance is to leave temperature, top_p and top_k untouched if the watermark matters for compliance.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
