Anthropic is adding its text watermark to four more Claude models. The mark is an imperceptible pattern in word choice. It adds no extra characters, tokens, or cost. It does not slow output speed. On Claude Opus 4.6 and Sonnet 4.6, developers can bypass it by tweaking one sampling parameter. Anthropic warns that relying on the watermark for compliance requires leaving those settings at default. The rollout meets the EU AI Act’s December deadline. Newer models like Opus 5.5 already carry the mark. The watermark works on every platform hosting these models. Customers cannot turn it off.
What changed
Watermark was only on models released after August 2, 2026.
What it unlocks
Bypassing the watermark on two older models by adjusting a single sampling parameter.
- 4 models added
- December 2026 EU AI Act deadline
- 3 sampling parameters
What you need to act on it
- default sampling parameters on Claude Opus 4.6 and Sonnet 4.6
- EU AI Act compliance requirement
Sources