Claude's invisible watermark reaches 4 older models Nov 3

From Tuesday, November 3, 2026, the output of four older Claude models carries an invisible text watermark. The models are Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. The date comes from an Anthropic email to API customers, reported by The Decoder and other outlets. Anthropic's own help page explains how the mark works, but as of October 31 it did not list November 3 in its model table. So we are writing the schedule as a reported date, not a date already shown in the public table.

The mark is not a visible stamp and not a hidden character. When the model picks its next word and several words would fit equally well, the choice is no longer left to ordinary randomness. A secret key steers the choice, together with the words before it. That pattern is what the detector reads. Anthropic says the method adds no tokens, changes no price, and does not change content or readability in its internal tests. The mark also stores nothing about the user, the organization, or the conversation.

Why November 3

The background is a European transparency rule. Article 50 of the EU AI Act requires providers of text-generating systems to mark their output in a machine-readable format that can be detected as generated by a Super Intelligence system. The obligation started applying on August 2, 2026. Models launched after that date, such as Opus 5.5, carry the mark from day one. Older models got a transition period, and this November 3 group lands about a month before the December 2, 2026 deadline that MACCUS cites from the Digital Omnibus. Anthropic applies the mark worldwide, not only for European users, because according to TechWeez it cannot yet scope it by region.

The caveats worth reading

Two models carry a specific condition. On Opus 4.6 and Sonnet 4.6, the mark only works when developers leave three sampling settings at their defaults: temperature, top_p, and top_k. Changing any one of them is enough to produce unmarked output, according to The Decoder's report on the same email. For Mythos 5 and Opus 4.7, the sources we read name no equivalent condition. That is not proof the other two are immune to other routes, but the documented limit so far covers only those two models.

The second limit sits on the checking side. Anthropic's detector is still in private preview, open to groups such as regulators, media, researchers, and fact-checkers. Ordinary users cannot check their own writing. A detection result is also a probability estimate, not a verdict. Anthropic itself says a positive result does not distinguish between Claude writing the text and heavily editing it, and a full rewrite can remove the mark. The absence of a mark does not prove a human wrote the text either.

Three models still have no specific date after this group: Opus 4.5, Sonnet 4.5, and Haiku 4.5. Sonnet 4.5 is scheduled to leave the Claude API on November 30, according to Anthropic's September 30 release notes as cited by MACCUS. How model claims are measured, and why a leaderboard order can change when the fields counted change, is in our GDPx article. Another new model released this week, built for coding work that runs long, is in our LongCat article.

Sources: Anthropic email to API customers (reported by The Decoder, Nov 2, 2026), TokenPost (Nov 2, 2026), MACCUS (Nov 2, 2026), Anthropic help page on content marking (updated Oct 28, 2026). Figures are snapshots at writing time, morning November 3, 2026 WIB, and can change.

Written by Mastechon

Writing about crypto and Super Intelligence: market news, new projects, Super Intelligence models, and early opportunities like airdrops and testnets. Every piece starts from the official announcement, with the costs worked out up front.