By Hamza Faisal
Anthropic is extending its hidden text watermark to four more Claude models, but on two of them the mark comes with a catch: change a single sampling setting and it disappears. The gap, reported by The Decoder from an email Anthropic sent to API customers, sits at the centre of a transparency push meant to make AI-generated text identifiable ahead of a European deadline.
The rollout covers Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Newer models such as Claude Opus 5.5 already carry the watermark, and the change on Amazon Bedrock may take a few extra days to arrive. Anthropic describes the mark as an imperceptible pattern in word choice: it adds no extra characters or tokens, does not change pricing or the request format, and, in the company's account, does not change the meaning, quality or readability of the text.
One setting is all it takes
On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark applies only when developers leave three sampling parameters at their default values. Those parameters — temperature, top_p and top_k — control how much randomness the model uses when choosing each word. Change any one of them, even slightly, and the output comes back without the watermark.
Anthropic is open about the limitation in its customer email. "If you rely on the watermark for your own compliance, keep these settings at their defaults," the company wrote. No such caveat was reported for Claude Mythos 5 or Claude Opus 4.7, which suggests those two models are marked regardless of sampling settings, although the email does not explicitly confirm that. The distinction matters because Opus 4.6 remains Anthropic's best writing model, ahead of even Opus 5.5, according to The Decoder.
The watermark is embedded at the model level, so it travels wherever Claude is offered, including through Amazon Web Services, Google Cloud and Microsoft Foundry, and customers cannot switch it off directly. It uses a variant of Google DeepMind's SynthID text method, and a detection API is in private preview for regulators, researchers, journalists and fact-checkers.
Why Anthropic is retrofitting now
The timing is driven by the European Union's AI Act. Anthropic has signed the Act's Code of Practice on transparency for AI-generated content, and every Claude model released after August 2, 2026 shipped with the watermark from day one. Older models are being brought into line in stages, with a December deadline now approaching.
What the gap means for compliance
For organisations that planned to treat the watermark as evidence of provenance, the message is to audit the details, not just the model name. Which Claude version is running, and whether temperature, top_p and top_k are all at their defaults, now determine whether an output is marked at all on Opus 4.6 and Sonnet 4.6.
Anthropic also lists limits that apply even when the mark is present. A detected watermark means text may have come from Claude, not that Claude wrote it unaided, since people use the model to proofread or translate. A missing mark proves nothing either, because short passages, heavy edits and paraphrasing can wipe it out. Whether watermarking leaves writing quality untouched is also disputed.
None of this makes the expansion pointless: coverage across the current Claude lineup is now far broader. But until the older models are fixed rather than merely warned about, the watermark carries an asterisk — reliable on the newest models, and on two older ones only for as long as nobody touches the dials.
