Anthropic tells API customers it will watermark text from four more Claude models from 16 November, The Decoder reports
The email covers Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, and says the mark applies only when temperature, top_p and top_k stay at defaults, so changing one setting avoids it.
Why it matters
Text watermarks are one of the few tools for telling AI-written text apart from human writing, and Anthropic says its mark is an imperceptible pattern in word choice that adds no tokens or user data. The Decoder reports that Opus 4.6, first due for the mark on 3 November, slipped to 16 November without a stated reason. The sampling loophole matters because anyone relying on the mark for compliance can lose it by adjusting one parameter. Watch Anthropic's detection API, now in private preview, and whether the caveat is removed.
Key facts
- Models: Claude Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5
- Start date: 16 November 2026
- Opus 4.6 delayed from 3 November
- Mark applies only with default temperature, top_p and top_k
- Detection API in private preview