Anthropic extends Claude's text watermark to older models, with a sampling caveat on two

Anthropic is extending the invisible watermark it embeds in Claude's text to several models released before August 2026. According to The Decoder, which saw an email the company sent to API customers, the mark is being added to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Models launched after August 2, 2026, such as Claude Opus 5.5, already carry it from release.

What changes for older models

The rollout happens in stages, and The Decoder says Amazon Bedrock may need a few extra days. The sources reviewed do not give a single confirmed start date; some reports mention November 3, which could not be verified here. At the time of writing, Anthropic's help page still listed Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 among models without text watermarks, while noting that the company is adding marks to outputs of models released before August 2, 2026.

The watermark works at the model level, so it appears whether Claude is used through Anthropic's own products or through cloud platforms such as AWS, Google Cloud and Microsoft Foundry, and customers cannot switch it off. Anthropic describes it as a subtle pattern in word choice that does not alter meaning or readability, adds no tokens and carries no information about users or their conversations. The Decoder reports the technique is a variant of Google DeepMind's SynthID text method, which nudges the randomness of word selection to leave a statistically detectable signal.

The sampling-parameter gap

The notable exception concerns Opus 4.6 and Sonnet 4.6. On those two models, the email says the mark is applied only when the temperature, top_p and top_k parameters keep their default values, so a small change to any of them is enough to avoid it. Anthropic advises customers who rely on the watermark for compliance to leave the settings untouched. The Decoder notes the email does not attach this caveat to Mythos 5 or Opus 4.7.

Regulation and detection

Anthropic ties the retrofit to the EU AI Act's transparency rules, citing a December deadline, according to The Decoder. Detection is still restricted: a detection API is in private preview for eligible groups such as regulators, media, researchers and some enterprises. Anthropic's help page also warns that a detected mark only suggests Claude was involved, and that a missing mark does not prove a human wrote the text, since editing, translation or very short passages can weaken the signal.