Starting November 3, 2026, Anthropic will apply its text watermark to four older Claude models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Anthropic announced the change to its API customers by email, as The Decoder reports. The marking applies worldwide across all Claude products, from the Claude app to Claude Code and Cowork to the API, and you cannot turn it off.
The first wave of the changeover covered Claude Fable 5, Sonnet 5, and Opus 4.8 at the end of September.
What changes on November 3?
From then on, the responses of the four older models will also contain a pattern in word choice that isn't noticeable when reading. Anthropic doesn't add any characters or tokens for this. Prices as well as request and response formats stay the same, and the marking contains no information about you, your organization, or your conversations.
Anthropic is switching over all platforms that offer the models, including Google Cloud and Microsoft Foundry. On Amazon Bedrock, the switch may take a few days longer.
When is the marking absent?
For Claude Opus 4.6 and Sonnet 4.6, Anthropic applies the watermark only as long as the parameters temperature, top_p, and top_k are set to their default values. With these three values, developers determine how random the model's word choice is. If an application deviates from them, the watermark may be missing from the output. For Mythos 5 and Opus 4.7, the email does not mention this restriction, according to The Decoder.
When Claude texts stay marked
Text watermark also in Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 from November 3, 2026
Marking remains
- Text copied and pasted
- Some edits
- Opus 4.6 and Sonnet 4.6 with default values
Marking may be missing
- Very short texts
- Heavily reworded or translated
- Opus 4.6 and Sonnet 4.6 with changed values for temperature, top_p, or top_k
A match only shows that Claude was involved; a missing match does not prove human work
Sources: The Decoder, Claude Help Center
Even with default values, the marking is not reliable proof. On its help page on labeling, Anthropic writes that the watermark is preserved when copying and survives some edits. Very short texts, heavy rewording, or translations, on the other hand, can make the watermark unrecognizable. A match only shows that Claude was involved in a text, and a missing match does not prove that a human wrote it.
How sensitive such methods are is shown by measurements that OpenAI has published for its own text watermark: if every tenth word in a passage of 400 tokens is replaced with a synonym, the detection rate drops from around 92 to 66 percent. With a quarter of the words replaced, 17 percent remain.
How do OpenAI and Google proceed?
OpenAI introduced its method textGrain on October 5, 2026, and intends to use it to label texts from ChatGPT and Codex only in the EU. In the API, textGrain remains off by default; developers can switch the marking on themselves worldwide. Google DeepMind has released SynthID Text as open source and built it into the Gemini models. Anthropic, by contrast, offers no switch: Claude marks in all products and regions.
You cannot check Claude texts yourself. Anthropic is releasing the detection API for now only in a closed test phase, for eligible organizations such as government agencies, research institutes, and media. The free Claude Content Checker only examines files such as images, videos, and audio up to 100 MB for Content Credentials, that is, embedded information about what a file was created with. The tool explicitly does not check text.
What this means for you
If you write client texts with Claude, it's best to assume that longer passages are marked. If a client prohibits the use of AI, clarify your way of working openly beforehand instead of relying on rewording. How to use Claude sensibly for drafts and revisions is shown in the Claude practical training on TutKit.
If you integrate Claude into your own tool via the API and rely on the watermark for labeling, it's best to leave temperature, top_p, and top_k unchanged for Opus 4.6 and Sonnet 4.6, as Anthropic advises. Which obligations under Article 50 of the AI Act apply to your product, you have to clarify yourself.
If you evaluate other people's texts, for example as a teacher or in an editorial office, a missing watermark is not suitable as proof of human work.
The Decoder reported on the email to API customers on October 31, 2026.
Sources
