📅 October 31, 2026 • The Decoder
Anthropic is extending its text watermarking system to four additional Claude models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Newer models in the family, such as Claude Opus 5.5, already carry the feature. The company describes the watermark as an imperceptible pattern in word choice, so generated text reads normally while carrying a statistical signal that can be checked later. The change applies through the API, and the company frames the feature as part of its transparency work.
How the watermark works
The mechanism does not add visible markers to the output. Instead, it subtly influences which words the model selects during generation, creating a distribution-level fingerprint that is invisible to a human reader but statistically detectable in analysis. Anthropic says text produced under the watermark can be attributed to a watermarked model without any noticeable change in style or quality for the end user. Because the signal lives in word-selection statistics rather than in the text itself, simple edits such as rephrasing single words do not necessarily remove it.
Where the loophole sits
On two of the newly covered models, Claude Opus 4.6 and Claude Sonnet 4.6, the watermark only applies when the sampling parameters stay at their defaults. These are temperature, top_p and top_k. Changing any single one of these parameters through the API is enough for a user to receive unmarked output. Anthropic has explicitly warned users who rely on the watermark for compliance purposes to keep the default settings in their applications.
Detection and regulation
The retrofit of older models is tied to transparency obligations under the EU AI Act, which push providers toward making AI-generated content identifiable. Alongside the rollout, Anthropic operates a watermark detection API in private preview. This service allows selected parties to check whether a given text was likely produced by a watermarked Claude model, though access is currently limited and not generally available. The approach mirrors transparency efforts elsewhere in the industry, where providers increasingly mark synthetic content to meet regulatory expectations.
Impact
With this expansion, watermark coverage becomes the standard across the current Claude lineup rather than a feature of only the newest models. The parameter loophole on Opus 4.6 and Sonnet 4.6 shows that enforcement is not yet unconditional, since a single non-default sampling setting disables marking on those two models. For teams building compliance-sensitive workflows on the API, keeping default sampling parameters becomes a functional requirement, and the private-preview detection API will determine how useful the signal is in everyday practice.
Sources: the-decoder.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)