Since 3 November 2026, text from Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark, joining newer Claude models. Nothing in your requests has to change, but Sonnet 4.6 calls with a custom temperature, top_p or top_k may not be marked, and the watermark does not settle your own EU AI Act duties.
What was released
The change was announced by Anthropic on 3 November 2026 in the Claude Platform release notes. It extends the text watermark to three Claude models that launched before 2 August 2026, the date from which, according to Anthropic, Claude models launched in the EU support machine-readable marking from day one.
Anthropic's help article, How Claude marks AI-generated content, keeps the current list. In summary:
- Marked everywhere, including cloud partners: Claude Fable 5.1 and 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5.
- Newly added: Mythos 5, Opus 4.7 and Sonnet 4.6. The rollout on Amazon Bedrock for these three completes by 10 November.
- First-party surfaces only: Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 are marked on Anthropic's own platforms, but the table shows no text watermark for them through AWS, Google Cloud or Microsoft Foundry.
The mark applies at the model level, so it appears in output from the Claude API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, worldwide. Files Claude generates, such as PNG or JPEG images, carry a signed Content Credential (C2PA) instead, where the platform offers Claude's file generation.
How the watermark works, and what it doesn't change
Anthropic's explainer, How Claude's text watermark works, describes a version of the SynthID-Text method published by Google DeepMind. When several next words are equally good, a secret key, rather than a plain random number, settles the choice. Anyone with the key can then estimate how likely it is that Claude wrote a passage.
According to the text watermark guide, the mark adds no tokens or hidden characters and carries no information about you or your organisation. Anthropic states it has a negligible effect on speed and does not change the price. Where only one answer is right, as in most code or a factual name, there is nothing for the mark to act on, so code is marked less than prose, and comments more than code.
What it means for teams building on Claude
Check your sampling settings if you rely on the mark. The guide says Sonnet 4.6 text generated with a non-default temperature, top_p or top_k might not carry the watermark, and the help article lists Opus 4.6 in the same limitation. If your product depends on Claude output being detectable, omit all three, including values a framework or client library sets for you. Because a framework can set these without you noticing, log one raw request from each service and check what it actually sends, not just your own config.
The watermark is a signal, not proof. Anthropic is clear that a detected mark means content may have been processed by Claude, not that Claude wrote it, and that a missing mark proves nothing. Short passages, heavy editing, translation, light proofreading and older models can all leave no detectable mark. Don't use it to accuse a user or a supplier of using AI, and don't use it as your only audit evidence of what an agent wrote. Keep your own logs for that, as we set out in what an AI agent audit trail contains.
Your own transparency duties still apply. Anthropic's help article tells teams that deploy Claude in their own products to assess what Article 50 of the EU AI Act requires of them. The watermark is Anthropic's marking of Claude's output. It does not tell a person that they are talking to an agent, or label AI-generated text published to inform the public. We covered those duties in when an AI agent must tell people it is an AI.
Plan for detection access. Watermark detection is in private preview. Anthropic offers it to regulators, media, researchers and similar groups, and to enterprises that need to check marks on their own Claude output to meet obligations under the Act, through a request form in the help article. If your compliance plan relies on checking marks, apply now rather than when an audit asks. Files are easier: the free Claude Content Checker reads Claude-issued Content Credentials today.
When not to change anything. If you call a covered model other than Sonnet 4.6 or Opus 4.6, the mark is already in your output whatever your sampling settings, and there is nothing to do. If you serve Opus 4.6 or older 4.5-series models through a cloud partner, the table shows no text watermark there, so record that gap in your compliance notes or plan a move to a covered model.
Frequently asked questions
Do I need to change my Claude API requests for the watermark?
No. Anthropic says the watermark needs no change to requests or response handling, adds no tokens or hidden characters, and does not change the price. The one exception is sampling. If you rely on the mark, omit temperature, top_p and top_k on Claude Sonnet 4.6, including values set by a framework or client library, because non-default values may leave the text unmarked. Opus 4.6 has the same limitation.
Does Claude's watermark work on Amazon Bedrock, Google Cloud and Microsoft Foundry?
For current models, yes. Anthropic's table shows text watermarks in cloud partner output for Claude Opus 4.8 and the 5-series Fable, Mythos, Opus, Sonnet and Haiku models. For Mythos 5, Opus 4.7 and Sonnet 4.6 the Bedrock rollout completes by 10 November. Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 are marked only on Anthropic's own platforms, not through cloud partners.
Can I check whether a piece of text was written by Claude?
Not yet, for most teams. Text watermark detection is in private preview for eligible groups such as regulators, media and researchers, and for enterprises that must check marks on their own Claude output under the EU AI Act. Anyone can register interest through Anthropic's form. Even with access, a detected mark only means Claude may have processed the text, and a missing mark does not mean a person wrote it.
Does the Claude watermark meet my EU AI Act obligations?
Not on its own. The watermark is Anthropic's marking of model output under the Code of Practice it signed. Anthropic tells teams that deploy Claude in their own products to assess independently what Article 50 requires of them. That can include telling people they are dealing with an AI system and labelling AI-generated text published on matters of public interest, which the watermark does not do for you.
