Anthropic Extends Its Text Watermark to Older Claude Models, With One Catch

Anthropic is rolling its invisible text watermark out to four more Claude models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. API customers were told by email. Newer models such as Opus 5.5 already carry the mark, and the rollout on Amazon Bedrock may take a few extra days.

First, what a text watermark is. It is not a visible label or a hidden character. Anthropic describes it as an “imperceptible pattern in word choice”: the model nudges which words it picks, in a way a detector can later spot statistically but a reader never notices. It adds no tokens (the small text chunks models count and bill by), changes no prices, barely slows output, and contains nothing about you or your chats. Reportedly it is a variant of Google DeepMind’s SynthID method for text.

Here is the catch. On Opus 4.6 and Sonnet 4.6 the watermark only applies when the sampling settings temperature, top_p and top_k stay at their defaults. These settings control how adventurous the word choice is, and changing even one of them avoids the mark. Anthropic’s own email says: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” The warning is missing for Mythos 5 and Opus 4.7, which suggests those are marked regardless, though that is The Decoder’s inference, not an Anthropic statement.

What’s behind it: The driver is regulation. Anthropic points to the EU AI Act’s December deadline for marking machine-generated content, and models released after August 2, 2026 already ship with the mark. Older models are being retrofitted in stages. Because the mark lives inside the model, it applies on AWS, Google Cloud and Microsoft Foundry too, and customers cannot switch it off. OpenAI has its own system, textGrain, which is on by default only for EU ChatGPT and Codex users, while Google has open-sourced SynthID for text.

There are fair objections. Commentator John Gruber argues that synonyms rarely mean exactly the same thing, so choosing words to carry a key rather than for best fit can cost precision. Artificial Lawyer points out that law firms could be exposed if a client has banned AI use and can detect the mark. Anthropic has a detection API, but it is still in private preview.

What this means for you: If you simply chat with Claude, nothing changes you can see. If you are a developer who needs your output to be detectably AI-generated, leave the sampling settings alone on the 4.6 models. If you are a student or writer wondering whether a detector could flag your text, remember that watermarks are statistical, can be weakened by heavy editing, and only a tool with access to the detector can check them. For most people, this is a quiet compliance step, not a feature.

Sources

• Anthropic: How Claude marks AI-generated content
• The Decoder: Anthropic’s watermark rollout leaves a loophole on two older Claude models
• The Decoder: OpenAI will watermark ChatGPT text in the EU