Starting November 3rd, Anthropic is extending its watermarking system to four older Claude models: Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. The company's newer models have already received this marking.
Starting November 3rd, Anthropic is extending its watermarking system to four older Claude models: Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. The company's newer models have already received this marking.
According to The Decoder, Anthropic notified customers of the changes via email.
The watermarks will be invisible to readers, but will allow special tools to determine whether Claude participated in the creation or editing of the text.
However, in two models, Opus 4.6 and Sonnet 4.6, the watermark can be removed by changing the default text generation settings. If you change any one of the three parameters — temperature, top_p, or top_k — that determine how random the model is in choosing words, the watermark will not appear. Anthropic has already warned customers who rely on the watermark for regulatory compliance about this.
Previously, dev.ua wrote about how Claude watermarks work, whether they can be removed by editing the text, and why the program code will receive less pronounced marking.
