The Claude text watermark is now live on several Claude models, and more are getting it on November 16. If you use Claude for writing or through the API, your output may carry a hidden statistical pattern you can’t see or turn off.
Here’s what Anthropic has confirmed, what is only reported so far, and what you should actually do about it.
Here are the facts worth knowing before you read on.
- Anthropic’s release notes say Claude Mythos 5, Opus 4.7 and Sonnet 4.6 carry the watermark from November 3, 2026.
- Claude Fable 5.1 and Mythos 5.1 have carried it since September 1, 2026.
- The Decoder reports four more older models join on November 16.
- The pattern is invisible and doesn’t change your prompts, pricing or request format.
That list is the short version. The details below matter if you rely on Claude output for work or compliance.
What the Claude text watermark actually is
Anthropic describes it as an “imperceptible pattern in word choice.” You won’t see extra characters or odd phrasing. The text reads the same as before.
According to The Decoder, it is a variant of Google DeepMind’s SynthID text method. That approach nudges the randomness in word choice so a detector can later spot a statistical signature.
It’s built into the model itself, so it applies on every platform that hosts it. Customers can’t switch it off, and Anthropic says it contains no information about you, your organization or your chats.
Which Claude models carry the watermark, and since when
The rollout is happening in stages, and not every date comes from Anthropic itself. This table separates what the company published from what has been reported.
| Model | Watermark status | Where it comes from |
|---|---|---|
| Claude Fable 5.1, Claude Mythos 5.1 | Watermarked since Sept 1, 2026 | Anthropic release notes |
| Claude Mythos 5, Opus 4.7, Sonnet 4.6 | Watermarked since Nov 3, 2026 | Anthropic release notes |
| Claude Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 | Reported start: Nov 16, 2026 | The Decoder |
| Claude Opus 5.5 | Reported to include it already | The Decoder |
Treat the November 16 date as reported, not confirmed. The Decoder also says Amazon Bedrock may take a few extra days.
The official Anthropic release notes list the September and November 3 changes.
The sampling-settings catch for developers
On Sonnet 4.6, Anthropic’s own note says text from a request with a non-default temperature, top_p or top_k might not carry the watermark.
The Decoder reports the same condition for the older models joining on November 16. Its quote from Anthropic is blunt: “If you rely on the watermark for your own compliance, keep these settings at their defaults.”
So if your app tweaks those three settings, your output may come back unmarked on those models. That matters for any team that expects every response to be detectable.
Why Anthropic is doing this now
The Decoder links the retrofit to the EU AI Act. It says Anthropic signed the Article 50(2) code of practice in July, and older systems get a transition period that ends around early December.
That timing explains why older models are being updated in waves rather than all at once. It’s a reported explanation, so check Anthropic’s own pages if your legal team needs the exact wording.
It also fits a broader pattern of tighter rules around Claude. We covered one earlier in our post on the Anthropic usage policy update.
What you should do if you use Claude
For everyday chatting, nothing changes. The watermark won’t alter your answers, and you can’t opt out.
If you publish Claude-written content, assume it could be flagged by a compatible detector. Edit it, add your own examples, and don’t pass raw output off as handwritten.
If you build on the API, test whether your sampling settings affect marking, and decide whether you need defaults for compliance. Anthropic’s detection tool is reportedly in private preview, so you likely can’t check it yourself yet.
One more point for teams: write down which models you use and which sampling settings you set. If a client or regulator asks whether your AI text is marked, you’ll want a clear answer, not a guess.
Schools and publishers may also start asking the same question, so a short internal note now saves a scramble later.
Frequently Asked Questions
Can I turn off the Claude text watermark?
No. Anthropic builds it into the model, and customers can’t disable it. On some models, non-default sampling settings may leave text unmarked.
Will the watermark change how Claude writes?
Anthropic says it’s imperceptible and adds no extra tokens or characters. It shouldn’t change meaning, readability, pricing or request formats.
Does the watermark show who wrote the text?
No. Per The Decoder, it contains no information about the user, their organization or their conversations with Claude.
Can anyone detect it today?
Not publicly. The Decoder reports that Anthropic’s detection API is in private preview, so most people can’t test text themselves yet.
Is the November 16 date official?
It’s reported by The Decoder, based on an Anthropic notice to API customers. The official release notes we checked list only the September 1 and November 3 changes.
My take: don’t panic, but stop assuming Claude output is anonymous. If your work depends on AI text passing as unmarked, that plan is getting weaker every month. If you’re building a product, lock your sampling settings to defaults where compliance matters, and be upfront with readers about how you use AI.
Leave a Reply