Anthropic is extending Claude's invisible text watermark to models that launched without it. Its help center says: "Consistent with our commitments under the Code, Anthropic is adding watermarks to outputs from models released before August 2, 2026." The Decoder reports that "In an email to API customers, Anthropic said it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6."

#Which models are affected?

Per Anthropic's help center, "Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch." For older models, it says "we’re working to add marking support for earlier Claude models."

The email names four older models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, starting November 3, 2026, according to The Decoder's German edition report of Anthropic's email to customers. On Amazon Bedrock, The Decoder reports, the email says "the rollout on Amazon Bedrock may take a few extra days." It also reports that "Newer models like Claude Opus 5.5 already carry the mark".

As of November 1, 2026, Anthropic's help page does not list Mythos 5, Opus 4.7, Opus 4.6 or Sonnet 4.6 as carrying text watermarks. Its table shows them with Content Credentials for files only. Claude Opus 4.5, Sonnet 4.5 and Haiku 4.5 have no text watermark listed either, and neither Anthropic nor The Decoder gives a plan for them.

Anthropic's release notes for September 1, 2026 say text from Claude Fable 5.1 and Claude Mythos 5.1 already "carries Anthropic’s text watermark," and that "Marking requires no changes to your requests or response handling."

#What is the caveat on two models?

The Decoder reports a limit on Opus 4.6 and Sonnet 4.6: according to its report of Anthropic's email, the watermark applies only when developers leave the model's sampling settings at their default values. The Decoder quotes Anthropic's email: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

The Decoder says the email does not mention that caveat for Mythos 5 or Opus 4.7. Its reading is that this "suggests the watermark applies to those models regardless of sampling settings." That is The Decoder's inference, not a statement from Anthropic. The Decoder also reports that "Customers can't turn it off."

No Anthropic page we checked mentions the sampling caveat.

#How does the watermark work?

Anthropic's help center says: "When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself." It adds: "You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response."

In its August 14, 2026 post, Anthropic explains the method: "Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick." It says the approach "is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024."

Anthropic makes these claims about it:

  • No extra cost. "Watermarking doesn’t require extra tokens, and will not be more expensive;"
  • No identity. "Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat;"
  • Everywhere. The help center says marks apply across "Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide."
  • Cloud partners. "When supported Claude models are accessed through AWS, Google Cloud, or Microsoft Foundry they will carry watermarks."

The watermark "will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing," the help center says. Anthropic's post is clear about the limits: "Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will." It also says detection "doesn’t work well on small samples" and that code "has generally less watermarking than some other forms of text."

Files are handled differently. "When Claude generates a supported file type such as a PNG or JPEG, it will attach signed provenance metadata," using Content Credentials (C2PA).

#Who can detect it?

"Watermark detection is currently in private preview, available to eligible organizations as required under EU law," the help center says, listing groups such as regulators, media, fact-checkers and independent researchers. Anthropic adds: "We plan to expand access to the detection API over time."

The free Claude Content Checker checks files only. "The tool does not check text," Anthropic says.

A detected mark is not proof on its own. Anthropic says it "tells you that the content may have been processed by Claude," and "It does not, on its own, confirm the full provenance of the content."

#Why is Anthropic doing this?

Anthropic gives the EU AI Act as its reason. The help center says: "Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems." It adds: "To support transparency and comply with our legal obligations, Anthropic is working to include machine-readable marks in content that Claude generates."

The Decoder, reporting on Anthropic's email, writes: "Anthropic points to the EU AI Act's December deadline as the reason it's watermarking older models now."

For more on Anthropic's models, see our coverage of Claude Mythos 5 and Georgia vote-count data and Claude Enterprise's smart reports and pooled budgets.

Our analysis: Extending the watermark to older models closes a gap for teams that stayed on Opus 4.6 or Sonnet 4.6, since their text was unmarked while newer models carried the mark. The reported caveat on those two models matters most for developers who change sampling settings, because their output may not carry the mark. Anthropic's help page had not been updated for the four models as of November 1, 2026, so the email, as The Decoder reports it, is the main public detail on timing and scope, including the November 3 date.

#What is unclear?

  • Timing. Anthropic's help center gives no date for the four models. The November 3, 2026 date comes only from The Decoder's report of the email, with no time or time zone.
  • Cloud partners. When Google Cloud and Microsoft Foundry output from the four models will carry the mark. The email, as The Decoder reports it, mentions only Bedrock.
  • Other older models. Whether and when Opus 4.5, Sonnet 4.5 and Haiku 4.5 get text watermarks.
  • The email itself. We have not seen it; The Decoder's report is the source for its contents.