The announcement was an email

Anthropic did not publish the November 3 rollout in a blog post. API customers got an email, reported by The Decoder on October 31, saying the company’s invisible text watermark would extend that Tuesday to four older models: Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Users of the consumer products got a similar message, reported by Mashable. The company’s public record of the change is a Help Center table, updated on November 3 — and the table does not match the email.

The table marks Mythos 5, Opus 4.7 and Sonnet 4.6 for text watermarking on Anthropic’s first-party surfaces and through cloud partners. Some cloud entries carry an additional footnote stating that the Amazon Bedrock rollout will complete by November 10. Opus 4.6 does not. Its row is blank in both text columns and ticked only for Content Credentials, the signed metadata Claude attaches to generated files. Opus 4.5, Sonnet 4.5 and Haiku 4.5 sit in the same unticked state, and none of those three was in the email. So the gap is specific: of the four models customers were told about, three appear in the public documentation and one does not.

The most plausible explanation is a documentation lag or staged rollout, not proof that Opus 4.6 output is unmarked. The help page does acknowledge that some cloud-partner rollout work remains unfinished: its Amazon Bedrock footnote gives November 10 as the completion date. But that does not explain why Opus 4.6 is absent from both text-watermark columns. Until Anthropic clarifies the discrepancy, the email and public table remain inconsistent. But a compliance officer working from the public page today — the page Anthropic points the world to — would conclude that Opus 4.6 text carries no watermark, while a customer who received the email would conclude that it does. Both cannot be the reference document.

The timing has a regulatory context. Article 50(2) of the EU AI Act establishes transparency requirements for AI-generated content. Anthropic signed the corresponding Code of Practice in July 2026. Its August explanation says models released before August 2 fall within a transition period and are being retrofitted over the following months.

One setting is enough

The email carries a second condition that the public page states more narrowly. On Opus 4.6 and Sonnet 4.6, according to the email, the watermark is applied only when three sampling settings — temperature, top_p and top_k, the dials that govern how randomly the model picks its next word — are left at their defaults. According to the customer email reported by The Decoder, changing any one of the three settings removes the watermark on Opus 4.6 and Sonnet 4.6. The public documentation describes the Sonnet 4.6 limitation less categorically, saying that outputs may not reliably carry the mark when any of those settings is non-default. Anthropic’s own sentence to customers is the bluntest version of this: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

The Help Center footnote covers the same ground with the edges sanded off. It names Sonnet 4.6 alone, and says outputs "may not reliably carry" the watermark at non-default settings. Read side by side, the two documents describe the same mechanism at two different strengths, for two different audiences: a condition in the email, a possibility in the footnote.

One possible explanation for the different treatment is the available sampling configuration. MadRobot, citing Anthropic’s API release notes, reports that Opus 4.7 and Opus 4.8 reject requests that attempt to use non-default temperature, top_p or top_k values. If that behavior is correctly described, those models cannot be tested under the alternative settings that the email identifies as a problem for the two 4.6 models. The email gives no caveat for Mythos 5 or Opus 4.7. On the two 4.6 models, which still accept custom sampling, coverage is a property of the request, not of the model.

What the mark can settle, and who can check

None of this makes the watermark itself weak. The method is a variant of Google DeepMind’s SynthID-Text. It uses a secret key to alter the source of randomness behind choices between near-equivalent words, creating a statistical pattern without inserting characters or extra tokens. Anthropic says the mark carries no identifying information about the user, organization or conversation, and can survive copying and some editing. Anthropic is unusually plain about what a positive result means: that Claude "was likely involved with the content at some point," not that it wrote the text rather than heavily editing it. Short passages, factual passages, proofreading and code carry little of the mark, because those are the places where word choice is not free. A complete rewrite removes it.

The sharper limit is access. Detection runs through an API in private preview. Anthropic says access is available to eligible organizations, including regulators, law enforcement, media, fact-checkers, researchers, educational organizations and EU civil society groups, as well as enterprises with comparable obligations under the AI Act. The free Claude Content Checker does not detect text watermarks; it reads Content Credentials embedded in files. There is no public page where a reader, an editor or an employer can paste a paragraph and test it. From November 3, a larger share of Claude’s output may carry a watermark that ordinary readers cannot independently check. Anthropic’s detection service remains in private preview, rather than being available as a public text-checking tool.

For anyone building on the 4.6 models with EU-facing output, the practical step is an audit: list every production call to Opus 4.6 and Sonnet 4.6, and record the temperature, top_p and top_k values it sends. Compare those values against the model’s documented defaults. If any differ, the customer email says the watermark will not be applied. Until Anthropic’s table and email agree about Opus 4.6, do not use the model name alone as evidence that a response is watermarked.

Documentation status: Anthropic’s Help Center table was last observed on November 4, 2026, following its November 3 update at 18:42 UTC. The Opus 4.6 text-watermark discrepancy is based on that version of the table.

Sources

  1. [1] Anthropic, "How Claude marks AI-generated content," Claude Help Center, updated November 3, 2026 (18:42 UTC) — re-read first-hand at Gate 1 on November 4, table unchanged from the drafting observation: Mythos 5, Opus 4.7 and Sonnet 4.6 ticked for text watermarks on first-party surfaces and through cloud partners, Opus 4.6 blank in both text-watermark columns and ticked only for Content Credentials in files; footnote: "Rollout on Amazon Bedrock will complete by November 10"; footnote for Sonnet 4.6 only: outputs "may not reliably carry the text watermark" when temperature, top_p, or top_k is set to a non-default value; marking applied at the model level on every surface; Anthropic has signed the EU AI Act Article 50(2) Code of Practice; detection in private preview for eligible organizations and obligated enterprises; the free Claude Content Checker reads file Content Credentials, not the text watermarkRead source
  2. [2] Anthropic, "How Claude’s text watermarking works," August 14, 2026, updated September 1 / October 14, 2026 (read first-hand at Gate 1 — the method is a version of Google DeepMind’s SynthID-Text (Nature, 2024); the mark changes the source of randomness used to pick among near-equivalent words, keyed by a secret key and the preceding words; "Nothing is added to the text and there are no hidden characters"; no extra tokens, no cost change; "Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat"; the limit, verbatim: "A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’"; detection works poorly on small samples; the mark is sparser on factual passages, proofreading and code; light editing probably will not remove it, "a complete rewrite where every word is replaced will"; applied globally because Anthropic does not yet have a durable way to scope it by region; the EU Code of Practice signed in July 2026 alongside around 190 signatories; models launched before August 2, 2026 in a transition period, being retrofitted "over the coming months")Read source
  3. [3] The Decoder, "Anthropic’s watermark rollout leaves a loophole on two older Claude models," October 31, 2026 (read first-hand at Gate 1 — the email to API customers: the text watermark would extend to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6; on Opus 4.6 and Sonnet 4.6 the watermark applies only when temperature, top_p and top_k are left at their default values; the advice, quoted verbatim: "If you rely on the watermark for your own compliance, keep these settings at their defaults"; no equivalent caveat for Mythos 5 or Opus 4.7; the EU AI Act December deadline as the stated reason)Read source
  4. [4] MadRobot, "Claude watermark comes to Opus 4.7 and Sonnet 4.6 on November 3," published November 1, updated November 3, 2026 (read first-hand in full at Gate 1 — Anthropic’s message to users, reported by Mashable, naming Tuesday, November 3 and the four models; the email’s defaults-only caveat via The Decoder; the API release notes point, as reported by MadRobot, that setting temperature, top_p or top_k to a non-default value returns an error on Opus 4.7 and Opus 4.8; the Bedrock rollout timing; the help page’s earlier state and its November 1 correction)Read source
  5. [5] ALFW, "Grok API Text Gets a Watermark by Default on November 7. Turning It Off Requires a Legal Attestation," November 1, 2026 — ALFW’s prior coverage of xAI’s watermark default, cross-referenced for the contrast in opt-out mechanics (Grok’s mark can be switched off with a legal attestation; on two Claude models it can be lost with a settings change) rather than as a source for new claimsRead source