THE CRUNCH
Anthropic is extending its text watermark to four more Claude models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Models such as Claude Opus 5.5 were already marked, and the change may arrive on Amazon Bedrock a few days later than elsewhere.
The catch sits with the two older models. On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark only applies when developers leave the three sampling parameters, temperature, top_p and top_k, which control how much randomness the model uses when picking words, at their defaults. Anthropic's email to API customers warns that anyone relying on the watermark for their own compliance should keep those settings untouched, since a minor tweak to any one of them is enough to dodge the mark. The email doesn't flag that caveat for Claude Mythos 5 or Claude Opus 4.7, implying those two are marked no matter how the settings are configured.
The loophole matters because Opus 4.6 is still Anthropic's strongest model for writing, outperforming even Opus 5.5. Anthropic points to the EU AI Act's December deadline as the reason for watermarking older models now: every model released after 2 August 2026 shipped with the mark on day one, while older models are getting it in stages. A detection API for verifying the watermark is in private preview.
The company describes the watermark as an imperceptible pattern in word choice that changes no meaning, quality or readability, adds no extra characters or tokens, and barely slows output. Because it is embedded at the model level, it works on every platform hosting these models, including AWS, Google Cloud and Microsoft Foundry, and customers can't turn it off. The method is a variant of Google DeepMind's SynthID, which tweaks word-selection randomness to produce a statistically detectable pattern. OpenAI's rival textGrain system, launched in October 2026, ships disabled and only activates by default for ChatGPT and Codex users in the EU.
Critics question whether the mark really leaves text quality alone. Blogger John Gruber contends that no two synonyms are exact matches, so choosing words by a watermark key rather than for meaning inevitably costs precision. Legal trade outlet Artificial Lawyer raises a separate worry for law firms: clients who have forbidden AI use might spot the mark, and documents carrying several watermarks at once could get messy.


