Estimated reading time: 5 minutes · Last updated:
Anthropic will extend its Claude watermark to four older models from 3 November 2026: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. The company told users at the end of October that these versions will get the same hidden marker already running in Claude Opus 5.5. Anthropic describes that marker as "an imperceptible pattern in word choice". The move answers pressure from EU legislators who want machine-written text marked so it can eventually be told apart from human writing. Until now, only Anthropic's latest models carried the mark. Anthropic says the change will not affect how accurate or readable Claude's answers are. It also says the pattern holds nothing that identifies the person or organisation using Claude.
Key takeaways
- Four more models: From 3 November 2026, Anthropic is extending its text watermark to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6.
- Same technique as Opus 5.5: The older models will use the watermark already running in Claude Opus 5.5, which Anthropic calls an imperceptible pattern in word choice.
- No user data in the mark: Anthropic says the Claude watermark does not lower accuracy or readability and carries nothing that identifies the user or organisation behind the text.
- Driven by EU demands: Anthropic adopted the watermark after EU legislators called for AI-generated text to carry a hidden mark separating it from human writing.
Table of contents
Which Claude Models Get the Watermark on 3 November
The first rollout covered only Anthropic's latest models, with Claude Opus 5.5 the version named as already carrying the mark. The expansion now reaches back across the lineup. From 3 November 2026, text produced by Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 will carry the marker as well.
Anthropic gave its users only a few days' notice. The alert went out on the Friday before the start date, and it presented 3 November as the point where the rollout begins. It did not say when coverage would be complete across all four models.
For teams that have pinned their work to one of these older versions, the practical effect is simple. Text they generate after the start date will carry the same marker as output from Opus 5.5. Staying on an earlier release will no longer keep output outside the watermarking scheme, at least for the four models Anthropic named.
How a Word-Choice Watermark Marks Text Without Changing It
A text watermark is a signal built into generated writing that software can look for later, even though a reader cannot see it. Anthropic's version works through vocabulary. The company says its mark is "an imperceptible pattern in word choice", so the signal is carried by the words Claude selects rather than by anything a reader would notice on the page.
Anthropic makes two promises about the technique. First, it says the pattern does not reduce the accuracy of answers or make them harder to read. Second, it says the mark contains no information identifying the user or the organisation that produced the text.
That second promise is the important one. A watermark that only says a machine wrote something is a labelling tool. One that also recorded who prompted it would be a tracking tool. On Anthropic's own description, the Claude watermark is the first kind. Whether outsiders will be able to check text for the pattern, and how, is not covered in the details Anthropic gave for this expansion.
Why Anthropic Is Watermarking Claude for European Lawmakers
The watermark exists because EU legislators asked for it. They want AI-generated text to carry a hidden mark so that, over time, writing produced by software can be separated from writing produced by people. Anthropic accepted that demand, as first reported by Mashable in early August 2026.
The aim is long-range. Marking a handful of models does not prove whether any given paragraph came from a person. It builds a signal that grows more useful as more machine-written text carries it, and adding older Claude versions widens that pool.
The pressure behind the policy is concrete. Lawyers have been caught filing AI-generated citations, novelists have been exposed for leaning on AI to write whole books, and researchers have been found using it in academic peer review. In each case the problem was a person passing off machine output as their own. What remains unsettled is whether broad watermarking will make that kind of dishonesty harder to get away with.
| Model | Watermark status | Start |
|---|---|---|
| Claude Opus 5.5 | Already active | Initial rollout |
| Claude Mythos 5 | Being added | 3 November 2026 |
| Claude Opus 4.7 | Being added | 3 November 2026 |
| Claude Opus 4.6 | Being added | 3 November 2026 |
| Claude Sonnet 4.6 | Being added | 3 November 2026 |
What the expansion could and could not achieve
The case for
- Coverage widens from Anthropic's latest models to four older ones, so more Claude output carries the same marker.
- Anthropic says the mark leaves accuracy and readability unchanged, which removes a reason for users to avoid watermarked models.
- Anthropic says the pattern carries no user or organisation data, which limits privacy objections.
The case against
- A mark is only useful if it can be checked, and Anthropic has not set out detection arrangements in this announcement.
- The watermark covers Claude only, so text from other providers is outside the scheme unless they adopt their own.
- The people with most reason to hide AI use also have the most reason to avoid or work around watermarked output.
What to be careful about
- Users who chose an older Claude version will find that output generated after 3 November carries the watermark, which may matter for workflows that publish text directly.
- Anthropic gave only a few days' notice before the start date, leaving little time for organisations to assess the change.
- The privacy assurance rests on Anthropic's own description of the mark, since no independent assessment of what the pattern contains has been published alongside the expansion.
The bottom line
The 3 November expansion is incremental but meaningful. It turns the Claude watermark from a feature of Anthropic's latest model into one that spans a wider slice of its lineup, including Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. What it leaves open is the harder question. A mark hidden in word choice only helps if somebody can check for it, and the people it is meant to catch have the most reason to avoid it. Anthropic's assurances on accuracy and privacy lower the cost for ordinary users. Whether the mark changes behaviour will depend on detection, and on whether European lawmakers press the same demand on every other provider.
What to watch
- Watch for the rollout to begin on 3 November 2026 across Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6; Anthropic has not given a date for completion.
- Watch for any tool or method that lets third parties check text for the Claude watermark; no date has been set.
- Watch for whether EU legislators extend the same demand to other AI providers; no date has been set.
Frequently asked questions
Does the Claude watermark reveal who wrote the prompt?
Anthropic says no. The company states that the mark holds no information identifying the user or organisation working with Claude, only a hidden pattern in word choice.
Which Claude models are watermarked from 3 November 2026?
Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 join the scheme from 3 November 2026. Claude Opus 5.5 already carried the watermark from the initial rollout.
Will watermarking make Claude's answers worse?
Anthropic says the word-choice pattern does not change the accuracy or readability of Claude's output. It is the same technique already in use on Claude Opus 5.5.
Related reading
