Claude now watermarks text from older models too, but on Opus 4.6 and Sonnet 4.6 changing temperature turns it off

If your content automations run on Claude, the text they produce is now carrying a hidden mark. On two older models, your settings decide whether it does.

What happened

Anthropic told API customers it will watermark output from Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, The Decoder reports. Newer models already carry the mark. Anthropic describes it as an imperceptible pattern in word choice that adds no tokens, does not change pricing and contains no information about the user. It is applied at the model level, works on AWS, Google Cloud and Microsoft Foundry, and customers cannot turn it off. Anthropic cites the EU AI Act's December deadline.

The catch: on Opus 4.6 and Sonnet 4.6, the watermark only applies when temperature, top_p and top_k are left at their defaults. Anthropic's email warns: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

Anthropic's help page adds that detection is in private preview for eligible organizations, that heavily edited or very short text may not carry a detectable mark, and that if you deploy Claude in your own product you should assess what Article 50 requires of you.

My take

For anyone running blog, email or social pipelines on Claude, three practical points:

1. Audit your nodes. Many n8n and Make flows set temperature by habit. On the two older models, that silently removes the mark.
2. Do not treat the watermark as your compliance. It is a signal, not a label. If a client sells into the EU, add a visible disclosure field in the CMS or post template and fill it automatically.
3. Log which model and settings produced each piece. When a client asks later, you can answer from data.

Also note OpenAI's own text watermark ships disabled for API users. Your pipeline may mix models with different rules, so track it per output.