Anthropic is adding text watermarks to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 to comply with the EU AI Act. The watermark uses a variant of Google DeepMind’s SynthID method to create a detectable pattern in word choice. The feature is embedded at the model level and cannot be turned off by customers on any hosting platform.
API users can bypass the watermark on Claude Opus 4.6 and Sonnet 4.6 by changing a single sampling parameter like temperature. Anthropic warns customers that relying on the watermark requires keeping sampling parameters at their default values. The company offers a detection API in private preview to verify the watermark on generated text.
Developers using these models must ensure their sampling parameters remain at default settings to maintain the watermark. The watermark is a technical compliance measure that affects how the model generates text. This change impacts developers who need to verify the origin of the content they generate or receive.
