Anthropic marks Claude's text with a hidden pattern, but two models have a gap

Anthropic is adding a hidden watermark to the text of four more Claude models. On two of them, a developer can switch it off by changing one setting. The two are Claude Opus 4.6 and Claude Sonnet 4.6. Anthropic told its API customers by email, and says the move is tied to the EU AI Act's December deadline.

What a text watermark is

A text watermark is a hidden pattern in word choice. Anthropic calls it imperceptible and says it does not change meaning, quality or readability. It adds no extra characters, does not affect pricing, and barely slows output. Anthropic also says it holds no information about the user, their organization or their conversations. The method is a variant of Google DeepMind's SynthID, which nudges the randomness behind word picks to leave a pattern software can detect.

What changed

• Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 now get the watermark.
• Newer models such as Claude Opus 5.5 already carry it.
• On Amazon Bedrock, the rollout may take a few extra days.
• Customers cannot turn the watermark off.

Every model released after Aug. 2, 2026 shipped with the watermark from day one. Older models are getting it in stages. Because it is built into the model itself, it works on every platform that hosts them, including Amazon Web Services, Google Cloud and Microsoft Foundry.

The loophole

On Opus 4.6 and Sonnet 4.6, the watermark only switches on when developers leave three settings at their defaults. They are called temperature, top_p and top_k, and they control how much randomness the model uses when picking words. Changing any one of them is enough to dodge the mark. Anthropic's only answer is a warning: if you rely on the watermark for your own compliance, keep these settings at their defaults.

The email does not give that warning for Mythos 5 or Opus 4.7, which suggests the mark applies there whatever the settings. The gap still matters, because Opus 4.6 remains Anthropic's best writing model, ahead of even Opus 5.5.

Open questions

Anthropic has a detection tool in private preview for checking the watermark. Whether the mark really leaves quality untouched is disputed. Blogger John Gruber argues no two synonyms mean exactly the same thing, so choosing words by a watermark key costs precision. The outlet Artificial Lawyer warns that law firms could face trouble if clients who banned AI discover the mark, or if several watermarks pile up in one document.

Rivals differ. OpenAI launched textGrain in October 2026. It is on by default only for ChatGPT and Codex users in the EU, and API customers can turn it on or off. It ships disabled.

What to do

If your business uses Claude through a developer tool and counts on the watermark for compliance, ask whoever built it to keep the three randomness settings at their defaults on Opus 4.6 and Sonnet 4.6.