Anthropic's Claude watermark expansion has a parameter bypass on two models

Anthropic is rolling out text watermarks to four additional Claude models to meet an EU AI Act December deadline, but the implementation has a gap: on Claude Opus 4.6 and Sonnet 4.6, API users can bypass the watermark by changing a single sampling parameter. The company addresses the loophole with only a compliance warning, raising questions about enforcement versus checkbox compliance.

• Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 are receiving the watermark expansion, with newer models like Claude Opus 5.5 already carrying it
• On Claude Opus 4.6 and Sonnet 4.6, Anthropic applies the watermark only at default settings, allowing developers to dodge it by tweaking one sampling parameter
• Anthropic describes the watermark as an imperceptible pattern in word choice that doesn't change meaning, quality, or readability, and contains no user or organization information
• The rollout on Amazon Bedrock may take a few extra days compared to the direct API
• Critics debate whether watermarking degrades text quality, with some arguing it sacrifices clarity while others contend it merely uses different random numbers during generation

Bottom line If you use Claude Opus 4.6 or Sonnet 4.6 via API and need watermark-free output for legitimate purposes, test your sampling parameters now before the December compliance deadline.