Anthropic extends Claude’s hidden watermark to four older models — but two have a loophole

Anthropic is extending its hidden text watermark to four older Claude models — Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 — from Tuesday, 3 November, according to an email the company sent to API customers. The move was reported on Saturday, 31 October. The watermark is an invisible pattern woven into the model’s choice of words, meant to show later whether Claude wrote a piece of text. Newer models such as Claude Opus 5.5 already carry it.

But the rollout has a catch. On Opus 4.6 and Sonnet 4.6, the watermark only appears when developers leave three sampling settings — temperature, top_p and top_k — at their default values. Those settings control how random the model’s word choices are, and changing any one is enough to dodge the mark. Anthropic warns customers who rely on the mark for compliance to keep the defaults. The email gives no such warning for Mythos 5 or Opus 4.7, suggesting those two are marked regardless.

Anthropic says the mark adds no extra characters, changes neither meaning nor readability, and carries no information about the user. It uses a version of Google DeepMind’s SynthID text method, and checking for the mark is limited to a private preview for regulators, researchers and journalists. The company is acting ahead of a December deadline under the EU’s AI Act, which requires AI-generated content to be marked so machines can read it.

Why it matters

The watermark is meant to make AI text traceable, so schools, publishers and regulators can tell when Claude wrote it. A loophole on two widely used older models weakens that promise: one changed setting produces unmarked text, and the public cannot check for the mark yet. OpenAI launched its own text watermark, textGrain, in October, but it is on by default only for ChatGPT and Codex users in the EU.

What happens next

From 3 November, text from the four models on the Claude Platform will carry the mark, with Amazon Bedrock following a few days later. Anthropic says access to its detection tool will widen, but it has given no date. The December EU deadline is likely to push other providers to publish their marking plans before then.

This article is for general information and education only. It is not legal or financial advice.

Sources

• https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/ (The Decoder, 31 October 2026)
• https://techprolonged.com/2026/10/anthropic-adds-text-watermark-to-four-older-claude-models/ (Tech Prolonged, 31 October 2026)
• https://pasqualepillitteri.it/en/news/27785/claude-watermark-november-3-opus-4-7-sonnet-4-6 (Pasquale Pillitteri, 28 October 2026)