Anthropic plans to add invisible text watermarks to output from four older Claude models starting November 3, according to an email sent to API customers and reported by The Decoder. The change would extend the company’s AI-content marking system to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6.
The customer notice says the rollout on Amazon Bedrock may follow by a few days. Anthropic’s public documentation says supported watermarks apply across Claude products and, where supported, through cloud partners worldwide. The company is adding marks to models released before August 2, 2026, as it works toward the European Union’s AI Act transparency requirements.
Anthropic describes its text watermark as an imperceptible pattern woven into a model’s word choices. The company says it adds no extra characters or tokens and does not change a response’s meaning, quality or readability.
The rollout has a technical caveat for two models. The Decoder reports that Claude Opus 4.6 and Claude Sonnet 4.6 will carry the watermark only when developers leave the temperature, top_p and top_k sampling parameters at their default settings. Changing any one of them can prevent the mark from being applied, according to the customer email. Anthropic advises developers who rely on the watermark for compliance to keep those settings unchanged.
Anthropic’s public model table, updated this week, still does not show text watermarks for these four models and does not list the November 3 date. The company says it is working to add marking to earlier models, so the specific schedule currently comes from the customer notice rather than the public table.
Anthropic cautions that a detected mark signals that content may have been generated or processed by Claude, but it is not conclusive proof of provenance. The absence of a detected mark does not establish that text was written by a person.
The expansion follows other efforts to make AI-generated content easier to identify. OpenAI has begun adding textGrain watermarks to ChatGPT and Codex, while Google has opened its SynthID Detector for checking supported AI-generated media.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
