Anthropic adds text watermarks to responses from four Claude models
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Machine-text labeling is becoming a built-in model feature, but on two older Claude models its presence depends on generation settings.
Anthropic told API customers it is rolling out a text watermark in Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The mark is a statistically recognizable pattern in word choice and adds no characters or tokens to a response. The company offers a detection API in private preview to verify authenticity. The staged addition of the mark to models released before August 2, 2026 is connected to the European AI Act's December requirements. On Opus 4.6 and Sonnet 4.6, the watermark is applied only at the default sampling settings. Changing temperature, top_p or top_k can prevent it from being formed, so Anthropic advises customers who rely on the mark for compliance to keep the defaults. The company's message stated no such condition for Mythos 5 or Opus 4.7.
Sources 1