Skip to content
Neural Daily

PolicyNews

Claude watermark: one setting disables it on two models

🇭🇺 Olvasd el magyarul →

Anthropic is extending its text watermark to four more Claude models ahead of the EU AI Act's December deadline. On Claude Opus 4.6 and Sonnet 4.6, though, the mark only applies at default sampling settings, so changing one parameter is enough to avoid it.

By: Mikola Máté Patrik · · 3 min read

Claude Help Center page titled How Claude marks AI-generated content
Image: Anthropic · Screenshot · Screenshot of the official Claude Help Center article · Source

What's changing

According to Anthropic's email to API customers, the watermark is coming to 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 it. The rollout on Amazon Bedrock may take a few extra days (The Decoder).

Anthropic calls the watermark an "imperceptible pattern in word choice." It adds no characters or tokens and does not affect pricing or request and response formats. It barely slows output and contains no information about the user or their conversations.

Where the loophole is

On Opus 4.6 and Sonnet 4.6, the watermark only applies when developers leave all three sampling parameters (temperature, top_p, top_k) at their defaults. Changing any one of them is enough to skip the mark. Anthropic's own email says that if you rely on the watermark for your own compliance, you should keep these settings at their defaults.

The email mentions no such condition for Mythos 5 or Opus 4.7, which suggests the mark applies there regardless of settings. The loophole stands out because, according to The Decoder, Opus 4.6 is still Anthropic's best writing model, ahead of Opus 5.5.

Why now

The driver is the EU AI Act's December deadline, which also covers systems already on the market. Anthropic's help page says every Claude model released after August 2 has been watermarked from day one, while older models are being retrofitted. The mark is applied at the model level, so it also appears in output on AWS, Google Cloud and Microsoft Foundry, and customers cannot turn it off.

The method is a variant of Google DeepMind's SynthID text technique. A detection API for verifying the mark is in private preview. OpenAI's textGrain, launched in October, is looser: it is on by default only for ChatGPT and Codex users in the EU, and ships disabled in the API.

What it means in practice

If your compliance depends on generated text being marked, leave the sampling parameters out of your Opus 4.6 and Sonnet 4.6 calls. And if you assume any Claude text is watermarked, these two models can prove you wrong. A missing watermark does not prove a human wrote the text.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles