Claude will start hiding a secret mark in text from older models

Anthropic will start hiding a secret mark in the text written by four older Claude models on November 3, 2026. The models are Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The mark is built into the words themselves, so copying and pasting the text does not remove it.

What changed

Anthropic is the company behind Claude. Newer Claude models, launched since August 2, 2026, already mark their text. Now the same mark is being added to models people have used for months. Anthropic emailed organizations that run at least one of the four models on the Claude Platform, the paid route where companies connect their own software to Claude. Amazon Bedrock customers may wait a few more days.

How the mark works

It is not a hidden character or a label attached to the file. When an AI writes, it picks each next word from a few that would fit equally well. The watermark makes that pick follow a pattern tied to a secret key. After "the sky was cold and," it might favor "gray" over "overcast." One choice proves nothing. Over a few hundred words, the pattern becomes statistically obvious to anyone holding the key.

Because the mark lives in word choice, deleting hidden characters does nothing. Anthropic says it can also survive some edits.

Who it affects

The biggest impact is on companies that stayed on the 4-series models through the API, out of cost or habit. Text they generate through November 2 goes out unmarked. After that, it carries the mark. Anthropic's own surfaces are covered too: the Claude Platform, the Claude app, Claude Code, Claude Cowork and Claude Tag.

The push comes from Europe. The EU AI Act requires makers of generative AI to mark synthetic content in a machine-readable way from August 2, 2026. Systems already on the market get until December 2, 2026. Anthropic applies the mark worldwide, not just in the EU.

What is not clear yet

• Opus 4.5, Sonnet 4.5 and Haiku 4.5 are not mentioned in the email, and Anthropic has given no date.
• Anthropic's support page shows a checkmark for all 16 models, but it describes the goal, not today's status.
• Anthropic's free Content Checker reads marks in files such as images, but it cannot detect the text watermark.
• Checking text is limited to organizations admitted to a preview.