Claude Text Watermark Extended to Mythos 5, Opus 4.7, Sonnet 4.6
Claude's text watermark now covers Mythos 5, Opus 4.7 and Sonnet 4.6. What the mark is, the Sonnet 4.6 sampling caveat, who can detect it and what to do.
PromptWises is reader-supported. Some links may earn us a commission at no extra cost to you. This never affects our ratings or recommendations.
Text written by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic’s invisible text watermark. The Claude Platform release notes added the change on November 3, 2026, with one caveat: on Sonnet 4.6, a request that sets temperature, top_p or top_k to a non-default value might not carry the mark. Nothing changes in how you call these models, and only a limited group of organisations can check for the watermark today.
What changed on November 3
Until now, Anthropic’s watermark covered newer models. Its help article, How Claude marks AI-generated content, says Claude models launched on or after August 2, 2026 support marking at launch, and that Anthropic is adding it to models released before that date. The November 3 note brings three of those older models in.
At the time of writing, the help article’s table shows this for text watermarks:
| Model | Claude apps and Claude Platform (API) | Cloud partners (AWS, Google Cloud, Microsoft Foundry) |
|---|---|---|
| Claude Mythos 5 | Yes | Yes, Amazon Bedrock rollout completes by November 10 |
| Claude Opus 4.7 | Yes | Yes, Amazon Bedrock rollout completes by November 10 |
| Claude Sonnet 4.6 | Yes, with the sampling caveat | Yes, Bedrock by November 10, with the sampling caveat |
| Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 | Not listed | Not listed |
The current models, including Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, were already ticked. The help article says marks apply across Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, worldwide.
What the watermark is, and what it is not
According to the text watermark docs, the mark:
- does not change the meaning, quality or readability of the output
- adds no tokens and no hidden characters
- carries no information about you or your organisation
- needs no changes to your requests or response handling
The help article explains that the watermark is woven into the text itself, so it travels with copied and pasted text and “may persist through some editing”. Anthropic’s August explainer on how the watermark works gives the background: only the source of randomness in word choice changes, and the mark cannot be traced to a specific person, organisation or chat.
Files are handled separately. Supported files Claude generates, such as PNG and JPEG, get signed C2PA Content Credentials, which anyone can check with Anthropic’s free Claude Content Checker.
The Sonnet 4.6 sampling caveat
This is the part developers should read twice. The docs say that on Sonnet 4.6, text generated for a request that sets temperature, top_p or top_k to a non-default value might not carry the watermark. If you rely on the mark, Anthropic says to omit all three parameters, “including any that a framework or client library sets for you”. Many SDK wrappers and agent frameworks set a temperature by default, so check your config rather than your own code alone.
Who can detect it
Watermark detection is in private preview. The help article lists eligible organisations as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises that need to check marks on their own Claude outputs for EU AI Act compliance. Others can register interest through the Claude Watermark Detector Access Request Form. Anthropic says it plans to expand access over time.
Anthropic is also clear about the limits. A detected mark is “not fully conclusive”: Claude may have only proofread or translated someone else’s text. A missing mark proves nothing either, because heavy editing, paraphrasing, translation, very short passages or older models can all leave no detectable signal.
What it means for you
- Developers: no code change is needed. If you build on Sonnet 4.6 and need marked output, drop custom sampling settings or move to a newer model. If you deploy Claude in your own product, the help article says to assess your own obligations under Article 50 of the EU AI Act.
- Teams publishing Claude-assisted text: assume text from supported models is marked, whichever Claude app or API produced it. Decide your disclosure policy on that basis, not on whether a reader can check it today.
- Educators: a detector exists, but most schools cannot use it yet, and Anthropic itself says a result is a signal, not proof.
The watermark does not change our view of Claude’s writing quality in our Claude review, and the help article says marks apply wherever Claude is offered, so none of the plans in our Claude pricing guide avoids it. For how to read vendor notes like this one, see how to read AI model announcements. More updates are in our news section.
Frequently asked questions
Does Claude watermark its text?
Yes, on supported models. Anthropic's help article lists text watermarks for Claude Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5, 5, 4.8 and 4.7, Sonnet 5.5, 5 and 4.6, and Haiku 5.5. Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 are not ticked for text watermarks in that table at the time of writing.
Can I turn off the Claude text watermark?
Anthropic's text watermark docs and help article do not describe an opt-out. The docs say you do not need to change your requests for it. The only documented way output may lack the mark is the Sonnet 4.6 caveat: requests that set temperature, top_p or top_k to non-default values might not carry it.
Can anyone check whether text was written by Claude?
Not yet. Watermark detection is in private preview for eligible organisations, such as regulators, law enforcement, media, fact-checkers, independent researchers and educational organisations, plus enterprises checking their own Claude outputs. Others can register interest through Anthropic's request form.
Does the watermark add hidden characters or personal data?
No. Anthropic's docs say the watermark adds no tokens or hidden characters, carries no information about you or your organisation, and does not change the meaning, quality or readability of the output.
