Anthropic Is Watermarking Four More Claude Models — With One Catch

 

Anthropic is expanding its invisible text watermark to four more Claude models on November 3, 2026 — Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 — according to a customer email reported on November 1. The move extends a marking system already running on newer models like Claude Opus 5.5, and it is tied to the European Union's AI Act transparency deadline in December.

What Is Changing on November 3

The four models join Anthropic's watermarking rollout in stages. The newer models already carry the mark; these four are older models released before August 2, 2026, which had a transition period under the EU rules. Rollout on the Claude Platform starts November 3, while Amazon Bedrock may follow a few days later, per the customer notice.

How the Invisible Watermark Works

Nothing visible changes. Anthropic's watermark does not add characters, a notice, or metadata you can see. Instead, it subtly changes the randomness behind the model's word choices — a statistical pattern woven into the text itself, similar in approach to Google DeepMind's SynthID-Text. That means the pattern can travel with copied text and may survive some editing. Anthropic says its tests found no effect on content, creativity, or readability. A detector estimates the likelihood that Claude helped produce a passage — it is not a yes-or-no proof, and detection access remains in private preview for regulators, media, researchers, educational institutions, and enterprises with compliance obligations. It carries no identifying information and cannot be traced to a specific person.

The Catch on the 4.6 Models

Here is the loophole worth knowing. On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark applies only at default generation settings. Change temperature, top_p, or top_k — the sampling controls that govern how random word selection is — and the mark may not be applied at all, according to the reported customer email. The same caveat was not specified for Mythos 5 or Opus 4.7. If you rely on watermarking for compliance, keep those settings at their defaults on the 4.6 models.

Why It Matters

This is part of a wider shift toward AI transparency. Publishers recently got a shared opt-out switch for AI training, and Google opened its SynthID Detector to the public — the industry is slowly building ways to tell what AI made and what it was trained on. But a watermark with a settings loophole is only as strong as its defaults, and a detector in private preview helps watchdogs more than everyday readers for now. The direction is right; the coverage is still catching up.

Short take: Invisible watermarks are a genuine step toward honest AI content — just don't treat them as foolproof yet. On the 4.6 models especially, one changed setting can quietly switch the mark off.