Anthropic reportedly plans to expand text watermarking to older Claude models, with a sampling loophole on two

Anthropic reportedly plans to extend its invisible word-choice watermark to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on November 3, driven by the EU AI Act. The mark is said to add no tokens and to be detectable only by regulators and researchers as of October 31. On the two 4.6 models, changing temperature or top-p reportedly prevents the watermark from applying.