On this page

On 2026-11-03, the Claude Platform release notes state that text generated by Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6 carries Anthropic’s text watermark on the platforms listed in *How Claude marks AI-generated content*. That line is a coverage expansion of an existing marking mechanism—not a claim that Anthropic invented text watermarks on that calendar day. The same notes flag a sampling caveat for Sonnet 4.6: when a request sets temperature, top_p, or top_k to a non-default value, the text output might not carry the watermark. DigIntoAI has not run watermark detection, not verified marks on those three models’ outputs, not probed sampling-parameter edges, and not audited Amazon Bedrock region-by-region rollout. Every operational detail below is drawn from the official release notes and Help Center as self-reported by Anthropic.

Related reading: How Do You Prove an API Didn't Swap Models? Anthropic Opens FlashProve, Claude Pricing (November 2026), and Anthropic: Unreleased Claude Model Finds Critical Real-World Vulnerabilities. FlashProve is a neighboring Anthropic trust／proof research note; pricing is the model／plan entry map; the cyber piece is a separate capability report. None of them is a Claude text-watermark coverage article.

What the November 3 note actually says

The November 3 bullet is short and specific. Generated text from the three named models carries the watermark where the Help Center’s marking article says platforms support it. It points readers to that Help Center page for the platform list and to the Text watermark documentation for detail. It does not announce a public, self-serve detector for every account. It does not say Bedrock already finished deploying the three models’ marks everywhere. And it does not erase earlier precedents: release notes for models such as Claude Fable 5.1 and Claude Mythos 5.1 (for example around 2026-09-01) already stated that generated text carries Anthropic’s text watermark. The useful frame for November 3 is therefore which additional call paths／model IDs are now called out as marked, not “text watermarking appeared for the first time.”

Where marks are supposed to apply

The Help Center article *How Claude marks AI-generated content* frames Anthropic’s commitments under the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. In that framing:

  • New models launched in the EU on or after 2026-08-02 are expected to support machine-readable marking at launch (embedded text watermarks; Content Credentials／C2PA on supported generated files).
  • Existing models released before that date are described as in progress—Anthropic is adding watermarks to earlier Claude models during a transition period.
  • Marks are described as applying across Claude Platform (API), Claude apps, Claude Code, Claude Cowork, Claude Tag, and—where supported—cloud-partner surfaces, worldwide, with the caveat that some platforms or features may not support every marking type.

Two complementary techniques appear in the Help Center: (1) imperceptible watermarks woven into generated text (travel with copy-paste; may survive some editing; applied at the model level so product surface alone does not remove them), and (2) C2PA Content Credentials on supported files Claude produces. This article’s search headline tracks the text-watermark coverage called out on November 3 for Mythos 5／Opus 4.7／Sonnet 4.6—not a full C2PA field guide.

Model table: three headlines inside a longer list

The Help Center’s “Which Claude models support watermarking” table includes Mythos 5, Opus 4.7, and Sonnet 4.6 among a broader roster (Fable／Mythos／Opus／Sonnet／Haiku lines and other listed IDs). The table distinguishes first-party surfaces from cloud-partner output (AWS, Google Cloud, Microsoft Foundry) and from C2PA file credentials. For readers of this piece, the operational takeaway is narrow: the November 3 release note names those three as carrying text watermarks now; the table shows they sit beside other already-listed models. Treating the search headline as “only these three models ever get marks” would misread the table.

Amazon Bedrock: through November 10—not “already fully open”

A Help Center footnote on the model table states that rollout on Amazon Bedrock will complete by November 10. Written against a 2026-11-04 reading of the sources, that means Bedrock deployment for the relevant marks is scheduled to finish by 2026-11-10—it is not safe to write that Mythos 5／Opus 4.7／Sonnet 4.6 watermarks are already fully live on every Bedrock path. DigIntoAI has not checked per-region Bedrock availability or mark presence on Bedrock responses.

Sonnet 4.6 sampling parameters can drop the mark

Both the November 3 release note and a Help Center footnote (†) align on the same edge: for Claude Sonnet 4.6, outputs may not reliably carry the text watermark when temperature, top_p, or top_k is set to a non-default value. The Help Center limitations section also notes that generation with non-default sampling on Opus 4.6 or Sonnet 4.6 can leave text without a detectable mark. For API callers who routinely override sampling, that is a deployment constraint: “the model is on the watermark table” is not the same claim as “every response from this model ID will detect as marked.” DigIntoAI has not measured how often non-default sampling actually strips detectability.

Detection stays private preview

Watermark detection is described as private preview. Eligible organizations named in the Help Center include regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups, plus enterprises that need to verify marks on their own Claude outputs for compliance obligations. Anthropic says it plans to expand access to the detection API over time and points to an interest／access request form. That is not a public detector open to every Claude account, and it is not “anyone can paste text into a free checker and get a definitive Claude／not-Claude verdict.” File-side Content Credentials have a separate free Claude Content Checker path in the Help Center; that path does not convert text-watermark detection into a mass-market public tool.

What a mark does—and does not—mean

The Help Center’s limitations section is the corrective against overreading:

  • A detected mark is a signal that content may have been generated or processed by Claude. It does not, on its own, confirm full provenance.
  • Claude may not be the original author—proofreading, translation, summarization, or format conversion can still leave a mark on text whose ideas came from elsewhere.
  • Marked content can be modified, excerpted, or mixed after Claude processed it.
  • Absence of a detected mark does not mean the content was not AI-generated or processed. Causes listed include models before marking support, heavy edit／paraphrase／translation／mixing, very short passages with weak signal, stripped file metadata (for C2PA), unsupported platform／feature／file type, and the Opus 4.6／Sonnet 4.6 non-default sampling case above.

In short: a watermark is not a finding of authorship fraud or academic cheating by itself, and no watermark is not a clean bill of “human-only.” DigIntoAI has not validated detection false-positive／false-negative rates.

Operator checklist (we haven't retested)

  1. Clock. Platform release notes November 3, 2026 name Mythos 5／Opus 4.7／Sonnet 4.6 text watermark coverage; read Asia/Shanghai local calendar alongside any UTC stamps on secondary mirrors.
  2. Narrative. Coverage extension of an existing mechanism—earlier models already had “generated text carries watermark” language.
  3. Scope. Help Center table lists those three and other models; search headline highlights this round’s named trio.
  4. Bedrock. Rollout completes by 2026-11-10—Bedrock is not already fully open for these marks.
  5. Sampling. Sonnet 4.6 non-default temperature／top_p／top_k → mark may be absent; Help Center also flags Opus 4.6／Sonnet 4.6 non-default sampling under limitations.
  6. Detection. Private preview for eligible organizations／compliance enterprises; expansion planned—not universal public lookup.
  7. Semantics. Mark ≠ sole proof of authorship or cheating; no mark ≠ proof of non-AI; short／edited text can miss.
  8. Stamp. Official release notes／Help Center self-report; we did not retest.

Bottom line

Anthropic’s 2026-11-03 Platform release notes extend the public call-out of text-watermark coverage to Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6, on top of a Help Center model table that already situates those IDs among other marked models. Detection remains a private preview for qualified organizations rather than a public free-for-all; Bedrock rollout is written to complete by 2026-11-10; Sonnet 4.6 non-default sampling can omit the mark. Those are provider statements, not DigIntoAI lab results—we did not retest detection quality, sampling boundaries, or Bedrock regional progress. Field notes can go through the site-owner contact in the footer.