Text written by Claude now carries a hidden watermark even when it comes from Anthropic’s older models. Since Nov. 16, the Claude watermark covers Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, on top of every model Anthropic released from Aug. 2, 2026. You cannot see it, and it does not change what Claude writes.
The short version
Anthropic is adding an invisible watermark to text from all of its current Claude models, including older ones, to meet the EU’s rules on AI-generated content.
- From Nov. 16, 2026, text from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carries the watermark on supported platforms.
- The watermark applies in the Claude apps, Claude Code, Claude Cowork and the API, worldwide.
- Anthropic says it adds no hidden characters and carries no information about you, your organization or your chats.
- Only approved groups such as regulators, fact-checkers, researchers and schools can use Anthropic’s detection tool for now.
- The downside: light edits probably will not remove it, but text Claude only proofread may carry little or no watermark at all.
The real shift: AI-written text is starting to carry a fingerprint by default. If you use Claude for school, work or writing you publish, it is worth knowing what that fingerprint can and cannot show.
How the Claude watermark works
Claude writes one word at a time. Often two words would work equally well, such as “overcast” or “grey” to describe the weather. Anthropic says the watermark uses those low-stakes choices to leave a pattern that readers cannot notice but that a detection tool with the right key can find. In Anthropic’s own explanation, “Nothing is added to the text and there are no hidden characters.”
Because the pattern is part of the wording, it travels with the text when you copy and paste it. Anthropic says it may survive some editing.
Which Claude models carry the watermark
According to Anthropic’s help page, every Claude model launched in the EU on or after Aug. 2, 2026 marks text from launch. The Nov. 16 update added Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. The help page also lists Sonnet 4.6 and Opus 3 as covered.
Files Claude creates, such as images, get a separate label called Content Credentials, also known as C2PA. That is a signed note in the file’s details saying Claude helped make it, not a hidden pattern.
What it means for students, writers and businesses
- It does not identify you. Anthropic says the watermark cannot be traced to a person, an organization or a chat.
- Proofreading leaves little trace. Anthropic says text Claude only lightly edits is mostly your own words, so there may be little for the watermark to attach to.
- Code carries less of it. Code usually has one correct answer, so there are fewer free word choices for the pattern.
- Detection is limited for now. Anthropic’s checker is in private preview for regulators, law enforcement, media, fact-checkers, researchers, schools and some businesses.
- Developers, check your settings. On the older models, Anthropic says text may not carry the watermark if a request changes the temperature, top_p or top_k settings.
Anthropic says it is doing this to comply with the EU AI Act, and that it applies the watermark worldwide because it does not yet have a reliable way to limit it by region. Google uses a similar system for its own AI content, which our guide to the SynthID Detector explains.
If you use Claude Fable 5.2, our guide to how to turn it on covers what it does and where to find it.




