Four older Claude models get a watermark on November 16

Anthropic will add its text watermark to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 on November 16, according to a November 4 email to API customers reported by The Decoder. On all four, the mark is applied only while temperature, top_p and top_k stay at their default values, so if your team counts on it for compliance, changing any one of those settings leaves the output unmarked.

Anthropic’s advice in its first email on the retrofit was direct: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” Apart from that condition, there is no switch. The watermark is applied at the model level, so it appears in the Claude apps, Claude Code, Cowork and the API, and in text generated through AWS, Google Cloud or Microsoft Foundry. Customers can’t turn it off. OpenAI’s textGrain, launched in October, works the other way for developers: it is on by default only for ChatGPT and Codex users in the EU, and API customers anywhere can turn it on or off, with the setting off by default.

Which models, and when

The retrofit comes in two waves. Since November 3, Claude Mythos 5, Opus 4.7 and Sonnet 4.6 have been marking their output, and Notebookcheck notes that Anthropic’s help page already lists all three as supported. Opus 4.6 was due the same day. Anthropic confirmed the move to November 16 but gave no reason. Opus 4.5, Sonnet 4.5 and Haiku 4.5 were not in the first announcement at all. On Amazon Bedrock each wave may take a few extra days, and every Claude model released after August 2, Opus 5.5 among them, has carried the mark since launch.

In the first wave, the default-settings condition covered two models, Opus 4.6 and Sonnet 4.6. Anthropic didn’t attach it to Mythos 5 or Opus 4.7, and The Decoder reads that as those two being marked whatever the sampling settings. Among the models already marked, Sonnet 4.6 is the one where a non-default temperature, top_p or top_k means no watermark. From November 16 that also applies to all four models in the new wave.

The condition matters most on Opus 4.6, which The Decoder calls Anthropic’s best writing model, ahead of even Opus 5.5.

Anthropic gives the EU AI Act’s December deadline as the reason for retrofitting older models now. It has signed the code of practice for Article 50(2), which calls for machine-readable marking of AI-generated content, and TNW reports that the marking applies globally, not only to European users.

What a detected mark shows

The method is a variant of Google DeepMind’s SynthID text. Where several next words fit about equally well, a key and the few preceding words settle the pick, replacing an ordinary random number generator, and a checker holding the key can then estimate how likely it is that Claude produced the text. Anthropic says this adds no characters or tokens, leaves pricing and request and response formats unchanged, barely slows output, and that the mark “contains no information about the user, their organization, or their conversations with Claude.” Google has open-sourced SynthID for text and built it into its Gemini models.

According to Anthropic’s help page, the mark travels with the text when it is copied and pasted and may survive some editing, but heavily rewritten, translated or very short passages may not carry a detectable one. A detected mark does not prove authorship: text that Claude only proofread or translated carries it too. A missing mark doesn’t rule out AI either, since output from older models and heavily edited text may carry none. Images are handled separately. PNG and JPEG files get Content Credentials based on the C2PA standard, which anyone can read with the free Claude Content Checker.

Is the text any worse?

Anthropic calls the watermark an “imperceptible pattern in word choice” that leaves meaning, quality and readability intact. Blogger John Gruber disputes that. He argues that no two synonyms mean exactly the same thing, so choosing words by a watermark key instead of by what fits the meaning inevitably costs precision.

The legal trade outlet Artificial Lawyer raises a separate problem for law firms. Clients who have banned AI use could discover the mark, and several watermarks layered inside one document could cause complications.

Checking for the mark is still limited. Detection is in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations and companies with their own obligations under the law, and access requires an application.