Anthropic is extending text watermarking to older Claude models, describing the mark as an imperceptible pattern in word choice. But for some models, that mark depends on developers leaving sampling settings unchanged. The rollout also includes a newly announced date for four models.
Watermarking is arriving in stages
Anthropic told API customers that Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 would receive watermarks. A later email set November 16, 2026, for Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5. The rollout on Amazon Bedrock may take a few extra days.
Claude Opus 4.6 had initially been expected to receive its watermark on November 3. Anthropic confirmed the move to November 16 but did not give a reason. The later email did not mention Claude Mythos 5, Claude Opus 4.7, or Claude Sonnet 4.6; the article says their rollout likely went ahead on November 3 as planned.
The later group shares a sampling restriction: watermarking applies only when temperature, top_p, and top_k are set to their defaults. In the first group, that condition applied to Claude Opus 4.6 and Claude Sonnet 4.6.
Default sampling settings matter
Sampling settings influence how much randomness a model uses when choosing words. Anthropic says that for Claude Opus 4.6 and Claude Sonnet 4.6, changing any one of temperature, top_p, or top_k is enough to avoid the watermark.
That caveat matters to customers who rely on the watermark for compliance. Anthropic’s email advises keeping the settings at their defaults in that case. The email did not state the same limitation for Claude Mythos 5 or Claude Opus 4.7, which suggests their watermarking applies regardless of those settings.
The mark is embedded at the model level, so it works across platforms that host the models, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers cannot turn it off directly. For models with the sampling restriction, however, a changed setting can mean the watermark is not applied.
What Anthropic says the mark changes
Anthropic describes its watermark as an “imperceptible pattern in word choice.” The company says it does not alter meaning, quality, or readability, adds no characters or tokens, and does not affect pricing or request and response formats. It also says output speed is barely affected and the mark contains no information about users, their organizations, or their conversations with Claude.
The watermark uses a variant of Google DeepMind’s SynthID text method. The method adjusts randomness in word selection to create a statistically detectable pattern. Anthropic has a detection API in private preview for checking whether text carries its watermark, and it provides more information on a content labeling page.
Detection brings questions about use and quality
The source article notes that claims about unchanged text quality are disputed. Blogger John Gruber argues that synonyms are not perfectly interchangeable, so choosing words partly to encode a watermark could reduce precision. That is a criticism of the trade-off Anthropic says its method avoids.
Artificial Lawyer raises a different concern for law firms: clients who have prohibited AI use might discover a watermark, and multiple watermarks within a document could cause complications. These concerns show why the details of detection and application matter to organizations deciding whether to use generated text.
Anthropic says the retrofit is connected to the EU AI Act’s December deadline. Models released after Aug. 2, 2026, shipped with watermarks from launch, while older models are being updated in stages. That makes the sampling caveat especially relevant: teams relying on content labeling need to know both which model they use and whether its watermark depends on default settings.