The extension and the models involved
Anthropic is extending its watermarking system for generated text to four additional models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. The Decoder reports this, citing a notice sent to API customers. The most recent models, such as Claude Opus 5.5, were already marked from the outset, while the extension on Amazon Bedrock may require a few extra days. The marking is embedded at the model level and remains active on all platforms hosting them, including Amazon Web Services, Google Cloud and Microsoft Foundry, with no option for customers to disable it.
According to Anthropic, this is an imperceptible pattern in word choice that does not alter meaning, quality or readability, does not add characters or tokens and does not affect prices or request formats. The company also claims that the watermark does not contain information about the user, their organization or their conversations. These are the company's statements, not independent verifications.
The flaw affecting older models
The critical point concerns Claude Opus 4.6 and Sonnet 4.6: on these models, the watermark applies only if developers keep the three sampling parameters temperature, top_p and top_k at their default values, which govern the degree of randomness in word selection. Even a minimal change to just one of them is enough to evade the marking. In the communication, Anthropic warns that those relying on the watermark for their compliance must leave these parameters at their default values, without indicating any other corrective measures.
The company does not mention the same condition for Mythos 5 and Opus 4.7, which suggests that on the latter the marking applies regardless of sampling. The flaw remains relevant nonetheless because, according to The Decoder, Opus 4.6 is still Anthropic's best writing model, ahead of Opus 5.5. Anthropic has a detection API in private preview for verifying the watermark.
Regulatory deadlines and competing alternatives
Anthropic cites the December deadline of the EU AI Act as the reason for the move. All models released after August 2, 2026, have been marked from day one, while earlier models receive the watermark in successive phases. The method used is a variant of Google DeepMind's SynthID system for text, which modifies randomness in word selection to produce a statistically detectable pattern. Google DeepMind has made SynthID open source for text and integrated it into Gemini models.
OpenAI launched its own textGrain system in October 2026, enabled by default only for ChatGPT and Codex users in the EU; API customers worldwide can turn it on or off, and it is delivered disabled. The difference is significant: Anthropic does not allow deactivation, while OpenAI leaves it optional for the API. It remains to be seen whether the watermark truly preserves text quality: blogger John Gruber argues that, since perfectly equivalent synonyms do not exist, choosing words based on a marking key reduces precision. The publication Artificial Lawyer reports a further risk for law firms, where clients that prohibit the use of AI could discover the marking, and the presence of multiple watermarks in a single document could create complications. These are reported opinions, not verified data.
