Anthropic’s Help Center page, updated Nov. 9, now lists text-watermark support for Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6. Notebookcheck had reported that marking for these models began Nov. 3. Anthropic’s current model-support table also lists C2PA file credentials for all three.
A Claude watermark is a signal of possible Claude involvement, not an authorship verdict.
How Claude’s text watermark works
Anthropic says its watermark is based on SynthID-Text, a method that adjusts the randomness behind selected word choices. When several words would fit about equally well, a secret key and preceding words help determine which one Claude selects. The pattern is statistical: Anthropic says it adds no hidden characters or extra tokens. The company’s explanation says the method uses low-stakes choices, where choosing one suitable word over another does not change the intended meaning.
That design also explains why a short passage or a light edit can be hard to detect. A brief sample contains fewer word choices for a detector to assess. If Claude only corrects grammar and punctuation, Anthropic says there may be too few changed words to produce a detectable signal. Heavier editing, paraphrasing, translation or mixing the text with other writing can also affect detection.
What a detection result can tell you
A positive text-watermark result can indicate that Claude generated or processed some of the text. It cannot establish that Claude originated the ideas or wrote the whole passage. Someone may, for example, use Claude to proofread or translate text they wrote themselves.
A missing mark does not prove human authorship either. Anthropic lists short passages, extensive editing, paraphrasing, translation and mixed text among the reasons a Claude mark might not be detected.
Text detection is in private preview for eligible organizations. Anthropic’s public Claude Content Checker checks supported C2PA credentials in files; it is not a public text-watermark detector. C2PA credentials are signed provenance metadata attached to supported files, such as PNGs or JPEGs. The file credential and text watermark are separate kinds of marks, and file metadata can be stripped when a file is converted or resaved.
Anthropic’s current table marks text watermarks for all three named models. For Claude Sonnet 4.6, it cautions that non-default temperature, top_p or top_k settings may prevent reliable marking.
What an open-source test can show
DevDrift’s English-language video reports translation and rewriting experiments using Google’s open-source SynthID implementation and the creator’s own key. The tests include translating text into German and back, and having a different model rewrite it. DevDrift says the setup did not use Anthropic’s production detector, so its results describe that separate test implementation.
