Anthropic is adding its text watermark to four older Claude models, but on two of them changing a sampling setting turns it off, The Decoder reports

Anthropic is extending its text watermark to older Claude models. Its help centre page on how Claude marks AI-generated content, updated this week, says the company is adding watermarks to outputs from models released before 2 August 2026, and lists Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 among the models with first-party text watermarks. The page gives no start date.

The Decoder reported on 31 October, citing an email Anthropic sent to API customers, that the reason is the EU AI Act's December deadline for marking AI-generated content. It also reported a gap: on Opus 4.6 and Sonnet 4.6, the watermark is applied only while the temperature, top_p and top_k sampling settings are left at their defaults. According to The Decoder, the email tells customers who rely on the watermark for their own compliance to keep those settings at their defaults, and says the rollout on Amazon Bedrock may take a few extra days.

Neither Anthropic's page nor The Decoder gives a start date for the older models.

Watermarks are one of the main tools regulators expect AI companies to use so that machine-written text can be identified. A watermark that switches off when a developer changes a common setting is weaker than one that is always on, which matters for companies that use these models to meet their own disclosure duties.

• Confirmed Anthropic says it is adding watermarks to outputs from models released before 2 August 2026, and lists Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 with first-party text watermarks. Claude help centre
• Reported On Opus 4.6 and Sonnet 4.6 the watermark applies only with default temperature, top_p and top_k, and the stated reason is the EU AI Act's December deadline, The Decoder reported from an Anthropic email to API customers. The Decoder