Claude's invisible text watermark extends to four older models
From 3 November, Opus 4.7, Opus 4.6, Sonnet 4.6 and Mythos 5 carry the same imperceptible mark Anthropic has used on new models since August.
Anthropic has started adding its invisible text watermark to four older Claude models, Opus 4.7, Opus 4.6, Sonnet 4.6 and Mythos 5, from 3 November. Until now, only models launched on or after 2 August carried it from day one.
The watermark is woven into word choices as the model generates, without changing meaning, quality or readability. Because it is part of the text itself, Anthropic says it travels when the output is copied elsewhere and can survive some editing. It covers output everywhere Claude is offered, including the API, Claude.ai, Claude Code and Claude Cowork, and applies worldwide rather than only in the EU. The legal basis is Anthropic’s signature on the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content.
Why it matters: if you build on the API with any of these four models, text your product generates from today onward carries a detectable, persistent mark, wherever your users are. Anything that assumed older-model output was indistinguishable from unwatermarked text, for provenance tooling, dataset curation or content-disclosure policies, needs rechecking.
The caveat: Anthropic is explicit that the mark isn’t fully conclusive. It can indicate Claude was involved but can’t say who was responsible or confirm whether the text was edited afterward.