Anthropic is extending text watermarking to four Claude models, with a caveat for developers using two of them through the API: changing a single sampling setting can prevent the mark from applying. That makes the rollout relevant not only to people checking whether text came from a model, but also to organizations that treat watermarking as part of a compliance process.
What the watermark does
Anthropic says the watermark is an “imperceptible pattern in word choice.” It is designed to be statistically detectable without adding characters or tokens, and the company says it does not change meaning, quality, or readability. Anthropic also says it does not affect pricing or request and response formats, and has little effect on output speed.
The company says the mark carries no information about the user, their organization, or their conversations with Claude. It is embedded at the model level, so it applies across platforms hosting the models, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers cannot switch it off.
Three settings determine whether two models are marked
Anthropic named Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 in an email to API customers. Claude Opus 5.5 already has the watermark. The rollout through Amazon Bedrock may take a few extra days.
For Claude Opus 4.6 and Claude Sonnet 4.6, the watermark is applied only when developers use the default values for temperature, top_p, and top_k. These sampling parameters influence how much randomness the model uses when choosing words. Change any one of the three, and the watermark can be avoided.
Anthropic’s email cautions: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” The email does not describe the same condition for Claude Mythos 5 or Claude Opus 4.7, suggesting their watermark may apply regardless of sampling settings.
This distinction matters for teams that want a consistent way to identify generated text. A watermark that depends on API configuration may not be present in every output from the two older models, even if developers have not deliberately set out to remove it. The settings therefore become part of the operational picture for anyone depending on detection.
A phased rollout for older models
Anthropic says it is adding watermarks to older models ahead of the EU AI Act’s December deadline. Models released after Aug. 2, 2026, shipped with watermarking from day one; older models are being updated in stages. Because the watermark is built into the model, the same approach carries across services that host it.
Anthropic offers further information on a content labeling page and has a detection API in private preview for checking the watermark. The company uses a variant of Google DeepMind’s SynthID text method, which adjusts randomness in word selection to create a pattern that can be detected statistically.
Detection brings questions about quality and use
Claims that a watermark leaves writing quality untouched are disputed. Blogger John Gruber argues that synonyms do not mean precisely the same thing, so steering word choice according to a watermark key could reduce precision. That concern focuses on the trade-off between detectable patterns and selecting the best-fitting word.
Artificial Lawyer raises a different issue for law firms: a client who has prohibited AI use might discover a watermark, while multiple watermarks in one document could cause complications. These concerns point to practical questions for organizations adopting detection, from how they interpret a mark to how they handle text that may contain more than one.
Other systems take different approaches. OpenAI launched textGrain in October 2026; it activates by default for ChatGPT and Codex users in the EU, while API customers worldwide can enable or disable it and it ships disabled. Google DeepMind open-sourced SynthID for text and built it into Gemini models. Anthropic’s rollout puts the focus on a different detail: for two Claude models, default API sampling settings determine whether the watermark appears.