Anthropic will start applying its invisible text watermark to four older Claude models on November 3, 2026: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, according to an email the company sent API customers, as quoted by CellCog; The Decoder reported the same four models. On Opus 4.6 and Sonnet 4.6, the email says, the mark applies only when the model's sampling settings stay at their defaults.
The change brings the four in line with newer models such as Claude Opus 5.5 before a December 2 deadline under the EU AI Act. Anthropic's help center page on marking, checked at 4:20 p.m. ET on October 31, still listed all four without text watermarks.
What changes on November 3
The watermark is a statistical pattern in the words Claude picks. Anthropic's announcement of August 14 says watermarking "uses the key and a few words that come before to settle what word the model should pick," and that "nothing is added to the text and there are no hidden characters."
The email, as The Decoder reported it, says the mark adds no tokens and changes neither pricing nor request and response formats. CellCog quotes it as saying "rollout on Amazon Bedrock may take a few additional days."
Anthropic's help page says watermarking is applied at the model level, including for models reached through AWS, Google Cloud and Microsoft Foundry. Once the four are added, 13 of the 16 models on that page will carry the text mark, by SaasBuzz's count.
Model | Oct. 31 | From Nov. 3 |
|---|---|---|
Mythos 5 | No | Yes |
Opus 4.7 | No | Yes |
Opus 4.6 | No | Default settings only |
Sonnet 4.6 | No | Default settings only |
Opus 4.5 | No | No date reported |
Sonnet 4.5 | No | No date reported |
Haiku 4.5 | No | No date reported |
Nine other models | Yes | Yes |
Source: Anthropic's help center, October 31, 2026; November 3 column from Anthropic's email as quoted by CellCog; The Decoder reported the same four models.
The sampling condition
For Opus 4.6 and Sonnet 4.6, The Decoder reports, the email ties the watermark to three settings left at their default values: temperature, top_p and top_k, which control how much randomness the model uses in picking words. The email warns: "If you rely on the watermark for your own compliance, keep these settings at their defaults."
The email, as quoted, carries no such caveat for Mythos 5, which Anthropic offers to Project Glasswing participants, or for Opus 4.7. On Opus 4.7 the question barely arises: Anthropic's API release notes say setting temperature, top_p or top_k to a non-default value already returns a 400 error on that model. That leaves Opus 4.6 and Sonnet 4.6 as the two models in the batch where, per the email, the mark depends on a setting the developer controls.
Who can detect it
Detection is not public. The help page says watermark detection is in private preview for "eligible organizations as required under EU law," such as regulators, media, fact-checkers and researchers, plus enterprises that must verify watermarking for their own compliance. Others can register interest through Anthropic's access request form. The free Claude Content Checker is offered for files carrying Content Credentials, not for text.
A detected mark is not proof of authorship. The help page says it means content "may have been processed by Claude," which can include proofreading or translating someone else's work, and that heavily edited or very short text may carry no detectable mark.
The EU deadline behind the date
The EU AI Act's transparency rules, Article 50, began to apply on August 2, 2026, according to the European Commission's AI Act Service Desk timeline. The same timeline gives providers of systems that generate synthetic content and "were already placed on the market before 02 Aug 2026" until December 2, 2026, to comply with Article 50(2), the duty to mark output in a machine-readable way. November 3 leaves 29 days before that date.
Anthropic's help page says models launched on or after August 2 carry marks from launch. Of the three older models with no reported date, Sonnet 4.5 is due to retire from the Claude API on November 30, two days before the deadline, according to the release notes.
What it means for teams shipping Claude output
For most integrations no code changes. What changes is what the text carries: a product that passes these models' words to its customers will, from November 3, pass on a mark that approved organizations can test for. Anthropic's help page leaves the deployer's own obligations to the deployer.
If you deploy Claude in your own product, you should independently assess what Article 50 requires of your products and services.Anthropic, "How Claude marks AI-generated content," help center
Not everyone accepts that the mark is free. The Decoder cites the blogger John Gruber arguing that picking words by watermark key rather than fit costs precision. Anthropic's announcement says that "in internal testing, we've seen no impact of watermarking on the content, level of creativity, or readability of Claude's text."
A short checklist for teams using the four models:
List which model IDs your code calls, and on which platform. Bedrock may lag the Claude API by a few days, the email says.
If your own compliance relies on the mark and you use Opus 4.6 or Sonnet 4.6, check whether any request sets temperature, top_p or top_k.
Do not assume the watermark covers your own Article 50 duties: Anthropic tells deployers to assess them for themselves.
If you need to verify text, apply for detection access; for generated files, use the Content Checker.
