According to an email from Anthropic to API customers seen by The Decoder: To meet the EU AI Act deadline in 12, Anthropic has added text watermarks to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6; models released after 8 2 (such as Opus 5.5) include watermarks upon launch. The watermark is a statistical pattern embedded in word choices; officials state it does not increase characters or tokens, does not change pricing or request formats, and cannot be disabled by users. The email specifies: Opus 4.6 and Sonnet 4.6 only include watermarks when the three sampling parameters—temperature, top_p, and top_k—are kept at their default values.
Confidence: High (95) · Event time 2026-10-31
Reasoning
- Anthropic Help Center's support list has been updated: Opus 4.7, 4.6, 4.5, Sonnet 4.6, 4.5, Haiku 4.5, and Mythos 5 currently include watermarks only in outputs from Anthropic's own channels (API, Claude app, Claude Code, etc.); the cloud partner column (AWS, Google Cloud, Microsoft Foundry) is not yet checked; newer models in the 5 series and Opus 4.8 have watermarks in both, and generated files also include C2PA content credentials.
- For developers relying on watermarks for compliance, the condition on older models is critical: changing any sampling parameter results in outputs without watermarks; Anthropic only reminds in the email, "If you rely on watermarks for compliance, please keep default settings."
- This is a vendor action under the transparency obligations of Article 50 of the EU AI Act; Anthropic has signed the relevant code of conduct; according to The Decoder, OpenAI's text watermark launched in 10 is disabled by default in the API, and Google's SynthID text watermark is open-source and built into Gemini.
Usage
- Claude API users needing to meet EU transparency requirements: Keep the three sampling parameters at their default values for Opus 4.6 and Sonnet 4.6. The watermark detection API is in internal preview, available only to regulatory, law enforcement, media, fact-checking, research, and educational institutions, as well as enterprises with compliance obligations; registration is available via the application form in the Help Center article "How Claude marks AI-generated content"; checking C2PA credentials in files can be done using the free Claude Content Checker.
Risks
- The claim that "changing sampling parameters removes the watermark" comes from a customer email cited by The Decoder; Anthropic's help pages do not state this condition, and the full email has not been made public; the claim about the 12 deadline also comes from this report.
- The official help page itself states: Watermarks may not be detectable in short texts, after significant rewriting, or after translation; detecting a watermark only indicates that the content may have been processed by Claude and does not prove that the entire text was generated by Claude.
- There is debate over whether watermarks affect word choice precision (The Decoder cited criticism from John Gruber); officials state it does not affect quality, but no evaluations have been published.
Open · Claude Help Center: How Claude marks AI-generated content