A reported Anthropic user email expands text watermarking to models including Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The two older models reportedly apply it only with default sampling settings, so not every response from those models can be assumed to carry a mark. Official guidance distinguishes statistical patterns in word selection from C2PA information used for media, and says text detection is available through a limited preview. Availability also needs to be checked for the relevant API or cloud route. A watermark does not authenticate an individual author or establish that the content is true.