The Claude logo. Image: Anthropic

Text written by four of Anthropic’s most widely used older Claude models will soon carry an invisible watermark. Anthropic told users on Friday that from Tuesday, November 3, the watermark will be added to the output of Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, Mashable reports. Anthropic’s help page on how Claude marks AI-generated content says it is “working to add marking support for earlier Claude models”, but its table doesn’t show text watermarks for them yet.

Until now the mark has covered Anthropic’s newer models: Claude Fable 5.1 and Fable 5, Mythos 5.1, Opus 5.5, Opus 5 and Opus 4.8, Sonnet 5.5 and Sonnet 5, and Haiku 5.5. Anthropic hasn’t published the November 3 date itself.

Which Claude models get the watermark

According to Anthropic’s message to users, the watermark reaches four older models from November 3: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6.

Anthropic’s help page hasn’t caught up yet. Its table still leaves the text-watermark columns blank for all of the older models, both on Anthropic’s own surfaces and through cloud partners (AWS, Google Cloud and Microsoft Foundry); for now it ticks them only for Content Credentials on files. Claude Opus 4.5, Sonnet 4.5 and Haiku 4.5 aren’t in Anthropic’s November 3 list. Sonnet 4.5 retires on November 30 anyway. Newer models such as Opus 5.5 are marked everywhere.

The watermark isn’t limited to Europe. Anthropic says it applies it worldwide “because we don’t yet have a durable way to scope it by region”.

How Claude’s watermark works

Anthropic explained the method in an August post. When Claude writes, it often picks between words that are equally good, such as “overcast” or “grey” after “The weather today was cold and”. Normally a random number settles the choice. With the watermark, a secret key and the few words before it settle it instead, which leaves a statistical pattern across a long passage that anyone holding the key can test for.

Claude’s watermark is a version of SynthID-Text, which Google DeepMind described in a 2024 Nature paper. Anthropic says “nothing is added to the text and there are no hidden characters”, that it costs no extra tokens, and that it “carries no identifying information and can’t be traced to a specific person, organization, or chat”. It says internal tests showed no effect on quality or readability.

Anthropic is doing this to meet the EU AI Act. It signed the EU’s Code of Practice on Transparency of AI-Generated Content in July, and the law gives models launched before August 2, 2026 a transition period, which is why older Claude models are being added later. OpenAI is taking a similar path: it has started hiding its own watermark in ChatGPT’s writing in Europe.

What the watermark can’t prove

Anthropic is frank about the limits. A detected watermark “can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’”

  • Rewriting: light editing probably won’t remove the mark, Anthropic says, but “a complete rewrite where every word is replaced will”.
  • Short and exact text: short passages give too little signal, and factual passages, light proofreading and code carry less of the mark because there are fewer free word choices.
  • No mark isn’t proof: text from a model before marking began, or one that was heavily edited, can be AI-written and still show nothing.

Who can check for it

Not the public. Anthropic’s detection API is in private preview for groups it is required to serve under EU law: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus companies that must check marking for their own compliance. Others can register interest. Anthropic’s free file checker reads Content Credentials on images and other files, not text.

That is a different approach from Google, which opened its SynthID Detector to everyone in October, and it means a teacher can’t paste an essay into a Claude checker today. Educational organisations are on the eligible list, though.

Update: a loophole on two of the older models

Anthropic’s email to API customers, reported by The Decoder, carries a catch. On Claude Opus 4.6 and Sonnet 4.6, it says, the watermark is applied only when developers leave the three sampling settings, temperature, top_p and top_k, at their default values. Change any one of them and the text comes back unmarked. The email tells customers: “If you rely on the watermark for your own compliance, keep these settings at their defaults.”

The email gives no such caveat for Mythos 5 or Opus 4.7. Anthropic’s API release notes show why the newer models are different: on Opus 4.7 and Opus 4.8, setting temperature, top_p or top_k to a non-default value returns an error, so developers can’t change them at all. The email also says the rollout on Amazon Bedrock may take a few extra days. Anthropic’s help page table still showed no text watermark for any older model on November 3.

Why it matters

Plenty of apps, businesses and writers still run on Opus 4.7, Opus 4.6 and Sonnet 4.6, and from Tuesday what those models write becomes checkable by the organisations Anthropic lets in. The watermark is a probability, not a verdict, so anyone using it to judge a student or a writer needs to remember what Anthropic itself says it can’t prove.

Correction, November 1: an earlier version said Anthropic’s help page already listed the four older models, plus Opus 4.5, Sonnet 4.5 and Haiku 4.5, as supporting text watermarks on its own surfaces. The help page doesn’t yet show text watermarks for any older model; the November 3 date and the four models come from Anthropic’s message to users.

Sources: Claude Help Center: How Claude marks AI-generated content; Anthropic: How Claude’s text watermarking works; Mashable (Anthropic’s message to users); The Decoder (Anthropic’s email to API customers); Claude API release notes; Nature: Scalable watermarking for identifying large language model outputs

Latest Anthropic news

More Anthropic news