Anthropic reportedly told customers in an email dated September 24, 2026, that Claude Fable 5, Claude Sonnet 5, and Claude Opus 4.8 are scheduled to receive text watermarks starting September 30. Anthropic describes the mark as a statistical pattern in word choices—not hidden characters—and says a positive detection can indicate Claude processing without proving who originated the ideas or wrote every sentence.
Three Claude models are reportedly scheduled for watermarking
The reported schedule names Claude Fable 5, Claude Sonnet 5, and Claude Opus 4.8, with a planned start date of September 30, 2026. The schedule was reported before that date; it does not establish that the rollout has taken place.
How Claude’s text watermark works
Anthropic describes its approach as a version of Google DeepMind’s SynthID-Text. Rather than inserting a visible stamp, hidden characters, or text metadata, it influences choices among plausible next tokens—the units a language model uses to generate text. Those choices create a statistical pattern that can be checked with the appropriate key.
The words themselves remain readable as ordinary text. Retyping an identical passage preserves its sequence of word choices, because the mark is not a copy-and-paste tag. Short passages, substantial editing, paraphrasing, or mixing text can make the pattern harder to detect.
What a detection can—and can’t—tell you
Anthropic says a positive result indicates that Claude may have generated or processed the text. It does not establish that Claude supplied the ideas, wrote every sentence, or identify a particular user, organization, or chat.
A missing signal also does not establish human authorship: short text, heavy editing, paraphrasing, and mixed passages can make a watermark harder to detect. Code generally receives less watermarking where exact output is required, although it may appear in discretionary text such as comments.
Which Claude models Anthropic lists as supported
As of September 26, 2026, Anthropic’s support table lists four models with text-watermark support on first-party surfaces: Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5. The reported expansion names Claude Fable 5, Claude Sonnet 5, and Claude Opus 4.8 separately.
Anthropic says marking applies at the model level across supported Claude products and wherever those models are offered worldwide. Support depends on the particular model and product surface.
Who can use the text detector?
Anthropic says text-watermark detection is in private preview for eligible organizations and enterprises with relevant compliance obligations; it is not generally available to the public.
File credentials are a separate mechanism. Supported generated files may carry signed C2PA Content Credentials as metadata, distinct from the statistical pattern used to mark text.
