Claude's text watermark is expanding. In an email to API customers on October 31, Anthropic said that starting November 3, output from Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 will also carry a text watermark. Until now, only models released after August 2, 2026 shipped with the watermark by default. Extending it to older models is routine under the EU AI Act's transparency requirements, but Anthropic also acknowledged that changing one sampling parameter can make the watermark disappear on two older models, weakening its value for compliance.
Anthropic describes the watermark as an “imperceptible pattern in word choice”: the text reads unchanged, it adds no characters or tokens, affects neither price nor format, and barely slows output. Because the mark is embedded at the model level, it appears whether text comes from the Claude app, the API or Claude Code, or from versions hosted on Amazon Web Services, Google Cloud or Microsoft Foundry, and customers cannot switch it off directly. The detection API is available only in private preview to organizations such as regulators, media outlets and researchers, so ordinary readers cannot check a passage themselves. OpenAI's textGrain is switched on by default only in the EU, a different approach; see OpenAI's text watermark approach in the EU.
One changed parameter, and the watermark is gone
The problem sits with the two older models. The email states that Claude Opus 4.6 and Claude Sonnet 4.6 are watermarked only when all three sampling parameters — temperature, top_p and top_k — stay at their default values; changing any one of them produces unmarked output. Anthropic warns that users who rely on the watermark for compliance should keep the default settings. The email mentions no such limitation for Mythos 5 or Opus 4.7, suggesting their watermarks are unaffected by sampling settings, and the rollout on Amazon Bedrock will arrive a few days later.
What a watermark can — and cannot — prove
A watermark is only useful for tracing afterwards: a detected mark can only show that text was possibly generated by Claude; it cannot prove who wrote it or that its content is true. The quality debate is equally unsettled: some argue that choosing words to satisfy a watermark key must sacrifice precision, while the legal industry worries about watermarks appearing in documents where clients have banned AI use. Anthropic's technique derives from Google DeepMind's SynthID text method, which leaves a statistically detectable pattern by adjusting the randomness behind word choice; how widely detection tools are opened up will decide how much difference it makes in practice. See how SynthID watermark detection is opening up.
After November 3, output from the four models will carry the watermark, and the transparency duties in Article 50 of the EU AI Act reach a key point in December. What to watch: whether Anthropic closes the parameter gap on older models, and when the watermark detection API moves from private preview to more verifiers.
