On Tuesday, November 3, text from four older Claude models, Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, starts carrying an invisible watermark. You will never see it, but anything those models write or edit for you can be checked later. Here is how it works, what it cannot prove, and the ten minute audit worth doing before you publish another AI assisted word.
What changes on November 3
Anthropic started watermarking text from its newest Claude models on August 2, 2026, the day the EU’s AI transparency rules kicked in. Anything released on or after that date carries the mark from launch. Older models were promised the same treatment over the coming months.
That promise is now landing. Mashable reports the expansion beginning November 3, and The Decoder names the four models from an email Anthropic sent to API customers. Newer models like Opus 5.5 already have the mark. The rollout on Amazon Bedrock may take a few extra days.
Because the mark is built into the model itself, it follows the model everywhere, including AWS, Google Cloud and Microsoft Foundry. According to The Decoder, customers cannot switch it off.
A quick sourcing note. The November 3 date and the model list come from that customer email, as reported by The Decoder. On November 2, one independent write up found that Anthropic’s own support page on content marking did not list them yet. It does state that Anthropic is adding watermarks to models released before August 2. So the direction is official. The exact date is reported, not published.
Why is this happening at all? It is compliance, not a product feature. Anthropic and about 190 signatories in total signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Anthropic applies the mark worldwide because it has no durable way to limit it by region. The Decoder ties the older model retrofit to the EU’s December deadline.
How the watermark works, minus the jargon
Every time Claude picks its next word, several words usually fit equally well. After the weather today was cold and, either overcast or grey works. Normally a random number settles the choice.
A watermark changes where that randomness comes from. Instead of an ordinary random number, Claude uses a secret key plus the few words just before. The output still looks random to you. But anyone holding the key can run the text through a check and ask whether the word choices match what Claude would have made.
Anthropic explains it with Monopoly. Imagine playing with the digits of pi instead of dice. Every move still feels random, but someone who knows pi can look at the game afterward and spot the pattern.
Here is what that means in practice:
- Nothing is added to your text. No hidden characters, no extra tokens, no extra cost.
- Because the pattern lives in the words themselves, it travels when you copy and paste.
- Anthropic says it found no effect on quality, creativity or readability in internal testing.
- The method is a version of SynthID-Text, which Google DeepMind published in Nature in 2024.
- Files like PNG, JPG and SVG get something different: a signed C2PA content credential in the metadata.
The Opus 4.6 and Sonnet 4.6 wrinkle
This is the part that got people talking. Per the customer email, Opus 4.6 and Sonnet 4.6 only get the watermark when three sampling settings, temperature, top_p and top_k, stay at their defaults. Change any one of them and the mark does not apply. Anthropic’s email tells developers to keep those settings at default if they rely on the watermark for their own compliance.
The email reportedly carries no such caveat for Mythos 5 or Opus 4.7. The Decoder reads that as the mark applying regardless of settings, but that is its inference, not a confirmation.
What this really means: the watermark is a compliance layer, not a lock. If you build on the API, your sampling settings now affect your paper trail. If you review other people’s work, a missing watermark proves nothing.
What the watermark cannot prove
Before anyone treats this as a lie detector, read the fine print. Anthropic is blunt about the limits.
- It shows Claude was likely involved, not that Claude wrote the text. Anthropic says it cannot tell Claude wrote this from Claude heavily edited this.
- Short passages are unreliable. Confidence grows with length.
- Factual lines with one correct answer and exact code leave little room for a mark. Proofreading leaves almost none.
- A translation done by Claude does carry it, since Claude chooses every word.
- Light editing probably will not remove it. Replacing every word will.
- It cannot detect other AI tools, and no watermark does not mean a human wrote it.
- It carries no information about you, your company or your chats.
One more catch. Only Anthropic’s key can read the mark, so third party AI detectors cannot see it. Anthropic’s detection API is in private preview for eligible organizations such as regulators, media, fact checkers, researchers and educators. Regular readers cannot test a paragraph today.
Critics also push back on the quality claim. The Decoder notes that blogger John Gruber argues no two synonyms mean exactly the same thing, so choosing words by key instead of by fit must cost some precision. Anthropic disagrees, and the evidence is still thin on both sides.
Claude is not the only one doing this
| Provider | What it does | Default? |
|---|---|---|
| Anthropic | Statistical text watermark, plus C2PA on files | On worldwide, no opt out reported |
| OpenAI | textGrain, announced October 5 | Rolling out in the EU for ChatGPT and Codex, opt in for the API |
| SynthID-Text in Gemini since 2024 | On for Gemini text | |
| xAI | Grok API text, per xAI’s docs, from November 7 | On, with a per model team opt out |
There is no shared standard. Each lab uses its own key, so a Google detector cannot read a Claude mark and the reverse.
Your ten minute AI content audit
Almost nothing changes in your daily chats. But if you write, teach, edit or publish for a living, a few small habits will save you an awkward conversation later. Here is the step by step version.
- List where Claude touches your words. Think of the Claude app, Claude Code, Cowork, and any tool or plugin built on Claude. Many apps run Claude underneath, so check the settings page or ask the vendor.
- Sort each use by how much Claude wrote. Drafted from scratch, rewrote, or only proofread. The more Claude writes, the stronger the mark.
- Pick a disclosure line and use it. Something plain works: I researched this myself and used an AI assistant for the first draft. I edited and fact checked every claim. One honest sentence beats a detection argument.
- Keep your paper trail. Save your outline, notes and version history. A watermark says Claude was involved. Your drafts show you were responsible.
- Check the rules you work under. Client contracts, school policies, publisher guidelines and employer rules may already say something about AI. Read them now, not after a question comes up.
- If you build on the API, document your settings. Note the model, and if you rely on the mark for compliance, keep sampling at defaults on Opus 4.6 and Sonnet 4.6. Expect Bedrock to lag by a few days.
- If you review other people’s work, treat the mark as a clue. It is a probability, not a verdict. Start with a conversation, not an accusation.
Quick answers
Will Claude’s writing read differently?
No. Anthropic says watermarked and unwatermarked text are indistinguishable to readers.
Does it cost more or run slower?
No. Anthropic says the impact on speed is negligible and there are no extra tokens, so pricing stays the same.
Can it be traced back to me?
No. Anthropic says nothing in the mark or its key identifies a person, company or chat.
Is this only for Europe?
No. Anthropic applies it globally.
Can I check a paragraph myself?
Not yet. Detection sits with Anthropic and eligible organizations through a private preview.
Sources
- How Claude’s text watermarking works, Anthropic
- How Claude marks AI-generated content, Claude Help Center
- Anthropic’s watermark rollout leaves a loophole on two older Claude models, The Decoder
- Anthropic expands text watermarking to older Claude versions beginning November 3, Mashable
- SynthID-Text paper, Nature
- Code of Practice on Transparency of AI-Generated Content, European Commission
- OpenAI will watermark ChatGPT text in the EU, Mixed
- Output Watermarking, xAI docs
Know a writer, teacher or founder who uses Claude every day? Send them this before Tuesday.
Leave a Reply