Claude watermarking reaches three older models, with limits

Anthropic expanded Claude text watermarking to Mythos 5, Opus 4.7 and Sonnet 4.6 in its November 3 release notes. The practical change is coverage of three existing models. Users should still check their hosting platform and sampling settings before assuming an output carries a detectable mark.

This is a follow-on to the worldwide marking program ByteForward covered in August, when support for older models was still being added. It should not be read as the first launch of Claude watermarking or as the release of a public, universal AI detector.

Which models and platforms are covered?

Anthropic’s current support matrix lists text watermarking for all three newly covered models on first-party surfaces and cloud partners. The footnote matters: Amazon Bedrock’s rollout for Mythos 5, Opus 4.7 and Sonnet 4.6 is scheduled to complete by November 10. As of November 5, that completion date remains ahead.

Coverage is worldwide and follows the supported model across Claude products. The matrix distinguishes text marks from C2PA Content Credentials attached to supported files. A checkmark for file credentials does not by itself establish text-watermark support for that model.

For teams documenting their deployment, the useful record is the exact model, platform, generation date and request configuration. A generic statement that an application uses Claude leaves out the details needed to interpret this rollout.

Sonnet 4.6 has a sampling caveat

Anthropic’s developer documentation warns that Sonnet 4.6 outputs might not carry the watermark when temperature, top_p or top_k is set to a non-default value. Its advice for applications relying on watermarking is to omit all three parameters, including values inserted by a framework or client library.

That makes the effective request worth checking, rather than only the code written directly by the application team. A wrapper that supplies its own defaults can change what the API receives. Teams should audit those settings before describing watermarking as a reliable part of their content workflow.

The documentation says ordinary request and response handling need no changes. Anthropic also says the mark adds no tokens or hidden characters and carries no information about the user or organization. Those are properties described by the provider, not independent test results from ByteForward.

An embedded pattern, not a visible label

In its August 14 technical explanation, Anthropic described a version of Google DeepMind’s SynthID-Text approach. It changes the randomness used when selecting among plausible words, leaving a statistical pattern that can be checked with a corresponding key. Readers do not see an extra badge or a string of invisible characters.

Anthropic says constrained factual passages and code leave fewer interchangeable choices, so their marking can be sparser. Light proofreading also gives the system less newly generated text to mark. These details make a detector’s result more limited than a simple authorship verdict.

The detector is still in private preview

Access to Anthropic’s text-watermark detector remains limited to eligible organizations, including media, researchers and educational institutions, and qualifying enterprises checking their own outputs. The company accepts expressions of interest and plans broader access. That differs from its free Content Checker for file credentials.

A detected watermark can support the conclusion that Claude was involved. It cannot distinguish original generation from substantial editing of someone else’s work. Conversely, a missing mark does not prove human authorship: another AI model, older output, short passages or extensive rewriting can all leave no usable Claude signal.

What builders should do now

Separate three questions in any internal checklist: whether the deployed model and platform support marking, whether the actual request preserves that support, and whether the organization can access an appropriate detector. None is answered merely by choosing a Claude model name.

For Bedrock users, November 10 is the documented rollout milestone to watch. For Sonnet 4.6 integrations, the immediate task is checking sampling parameters. For editors and reviewers, the standing rule is simpler: treat a watermark as supporting evidence about processing history, alongside source records and editorial review, rather than a substitute for them.

Featured image: AI-generated editorial illustration with the official Claude logo, credited to Anthropic. The lens is a visual metaphor, not a detection tool.