Claude's invisible watermark hits more models Tuesday. You still can't check it.

Starting November 3, 2026, Anthropic expands Claude's invisible text watermark to four more models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. That date and model list come from an Anthropic email to API customers, as reported by The Decoder on October 31. Newer models such as Claude Opus 5.5 already carry the mark. Bedrock may lag a few days.

What the mark is. Anthropic's published explainer says Claude uses a version of Google DeepMind's SynthID-Text: it changes the source of randomness when the model picks among near-equally good next words. Nothing is added to the text—no hidden characters, no extra tokens, no price change. Readers cannot see the difference. The mark carries no user, organization, or chat identifiers.

Who can check it. Detection is a private preview. Eligible organizations under EU rules (regulators, law enforcement, media, fact-checkers, researchers, certain enterprises) can request access. Regular users cannot run the detector on their own essay. Anthropic says a hit only means Claude may have processed the text (including proofreading or translation), and a miss proves nothing—especially on short passages or heavy paraphrases.

The loophole. Per The Decoder's reading of the customer email: on Opus 4.6 and Sonnet 4.6, the watermark applies only when temperature, top_p, and top_k stay at defaults. Changing any one can dodge it. Anthropic's email reportedly warns compliance users to keep defaults. The Decoder did not report that caveat for Mythos 5 or Opus 4.7.

Why now. Models launched on or after August 2, 2026 already marked output at launch (EU AI Act / Code of Practice). Older models are being retrofitted. Anthropic applies the mark worldwide because it says it cannot yet scope by region. Marks apply across Claude apps, Claude Code, the API, and cloud partners when the model supports them.

Caveats. We have not seen the customer email ourselves; the November 3 list and the 4.6 sampling caveat are attributed to The Decoder. Watermarking is not proof of authorship. Not tested by MODEL. Not sponsored; no affiliate links.

Sources: Anthropic watermark explainer; Claude Help Center marking page; The Decoder (October 31); The Verge (August 17).

MODEL Editors