Claude’s Hidden Text Watermark Is Set to Reach More Models

A hidden mark is coming to more Claude models

Updated November 2, 2026, 12:41 AM Central

Anthropic is preparing to extend its invisible text watermark to four older Claude models, according to reporting on an email sent to customers. The planned expansion would bring more AI-generated writing under a system designed to leave a detectable statistical signal inside the words themselves.

Up to Speed with AI, citing the customer email, reported that the change is scheduled to begin November 3 for Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on Anthropic’s API platform. It reported that the Amazon Bedrock rollout could follow a few days later. The email’s schedule is a reported plan, not proof the switch has already happened.

Anthropic’s model coverage table says it is adding marks to models released before August 2, 2026. At the time we checked, it did not yet list text watermarking for those four models. That distinction matters for anyone trying to work out whether a particular response has been marked.

What an invisible watermark actually does

This is not a faint logo stamped on a page. Anthropic says its watermark works by subtly shaping the choices made as Claude generates text, creating a statistical pattern that a detection tool can look for. The output remains ordinary text, so the signal can travel with copied words.

But the mark is not a magic truth detector. Anthropic says detection is probabilistic: a positive result can indicate that supported Claude output was involved, but it does not prove who wrote the original material or whether a person used AI to draft, translate or edit it. A missing mark does not establish that a passage is human-written.

Short passages, heavy rewriting, translation and other changes can weaken or eliminate the signal. A quoted human passage included in Claude’s response could also complicate interpretation. That makes a watermark a clue, not a verdict.

Why Anthropic is making the change

The company says it is implementing the system in response to the European Union’s AI transparency framework and applying its approach globally. Anthropic says it is working with access-controlled detection tools rather than offering a universal public detector. Its technical explanation describes privacy safeguards and limitations.

For publishers, schools and businesses, the distinction is practical: detecting that a model processed text is not the same as discovering why it was used. A person might ask an assistant to polish a sentence they wrote themselves, and the resulting output could carry a mark.

The planned November rollout does not mean every Claude model or third-party deployment changes on exactly the same schedule. Anthropic’s coverage table and the individual provider’s documentation are the places to check as support changes.

How the rollout developed

August 2, 2026Anthropic identifies the date as the dividing line for newer models covered by its watermark approach. Anthropic
August 2026The company published its explanation of text watermarking and its limitations. Anthropic
November 3, 2026Planned start for four older models on Anthropic’s platform, according to reporting on a customer email. Up to Speed with AI