Anthropic watermarks four more Claude models, but sampling tweaks defeat it on Opus 4.6 and Sonnet 4.6

Anthropic is expanding its text watermark to four more Claude models, but on two of them API users can avoid the mark by changing a single sampling parameter, according to a report by The Decoder.

In an email to API customers, Anthropic said it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Newer models such as Claude Opus 5.5 already carry the mark. The rollout on Amazon Bedrock may take a few extra days.

The sampling loophole

On Claude Opus 4.6 and Claude Sonnet 4.6, Anthropic says the watermark applies only when developers leave three sampling parameters at their default values: temperature, top_p, and top_k. These control how much randomness the model uses when choosing words. A minor change to any one of them is enough to avoid the watermark.

Anthropic's email warns: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

The email does not repeat that caveat for Claude Mythos 5 or Claude Opus 4.7. The Decoder reads this as a sign that the watermark applies on those models regardless of sampling settings. The outlet also notes that Opus 4.6 remains Anthropic's best writing model, ahead of even Opus 5.5.

What Anthropic says the watermark does

Anthropic describes the watermark as an "imperceptible pattern in word choice" that does not change meaning, quality, or readability. According to the company:

• It adds no extra characters or tokens.
• It does not affect pricing or request and response formats.
• It barely slows output speed.
• It "contains no information about the user, their organization, or their conversations with Claude."

The watermark is embedded at the model level, so it works on every platform that hosts these models, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers cannot turn it off. Anthropic uses a variant of Google DeepMind's SynthID text method, which adjusts the randomness behind word selection to create a statistically detectable pattern.

Anthropic has a detection API in private preview for verifying the watermark, and further details are on its content labeling information page.

Why now

Anthropic points to the EU AI Act's December deadline as the reason for watermarking older models now. Every model released after Aug. 2, 2026 shipped with the watermark on day one, while older models are being covered in stages.

How other providers compare

OpenAI launched its own watermarking system, textGrain, in October 2026. It activates by default only for ChatGPT and Codex users in the EU. API customers worldwide can switch it on or off, and it ships disabled. Google DeepMind open-sourced SynthID for text and built it into the Gemini models.

Quality concerns are disputed

Whether watermarking leaves text quality untouched is contested. Blogger John Gruber argues that no two synonyms mean exactly the same thing, so choosing words by a watermark key instead of semantic fit inevitably degrades precision. The legal trade outlet Artificial Lawyer raises a separate concern for law firms: clients who have banned AI use could discover the mark, and multiple watermarks layered inside one document could create complications.

What this means

For developers who depend on the watermark for compliance, the 4.6 models carry a configuration requirement: any change to temperature, top_p, or top_k means the mark may not be present. Teams running Opus 4.6 or Sonnet 4.6 with custom sampling should check their settings against Anthropic's guidance.

The watermark is mandatory and set at the model level for Anthropic's covered models, while OpenAI's API version is optional and off by default. The detection API is still in private preview, so independent verification is limited to approved users for now. Anthropic's claim that quality is unaffected sits against the objections from Gruber and Artificial Lawyer, and the source reports no independent measurement settling the question.