Claude's invisible text watermark now covers three older models as the EU's December deadline nears

Watermarking AI text has been the easy part of the EU's transparency rules to announce and the hard part to actually ship. Anthropic started marking text from its newest models in August, when Article 50 of the EU AI Act took effect. The older models — the ones already deployed across thousands of products — were given a transition period, and Anthropic has been working through them since. November 3 was the next step: three more models over the line, with the rollout on Amazon Bedrock finishing by November 10.

What changed on November 3#

The November 3 entry in Anthropic's Claude Platform release notes is one sentence long: text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on the platforms its help center lists. Those platforms are broad — the Claude API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, plus AWS, Google Cloud and Microsoft Foundry.

There is a wrinkle worth noting. An email Anthropic sent to Claude Platform customers ahead of the change named four models, adding Claude Opus 4.6 to the list. The release note does not include it, and the help center's model table still shows no text watermark for Opus 4.6. The same table shows no mark yet for Opus 4.5, Sonnet 4.5 or Haiku 4.5 either. Anthropic says it is adding watermarks to models released before August 2, 2026 over the coming months — so Opus 4.6 looks like a model whose marking slipped, not one that was cancelled.

How the mark works — and what it survives#

The watermark is not a stamp on the text. It is a statistical pattern in the words Claude chooses, set by a secret key — a version of SynthID-Text, the method Google DeepMind published in Nature in 2024. Nothing is inserted into the output: no hidden characters, no metadata, no extra tokens. That is why it costs nothing and, Anthropic says, does not change the meaning or quality of the text in any way a reader could notice. It also carries no information about the person, organization or chat behind the request.

What the mark survives matters as much as what it is. It travels with the text when copied and pasted elsewhere, and may persist through light editing. A full rewrite removes it, and proofreading or very short passages leave too little signal to detect. In other words, it is a provenance signal for passages that stay close to what the model wrote — not a forensic tool for text that has been thoroughly reworked.

And that check is the second gate. Detection runs through an API in private preview, open to the kinds of organization the EU law names: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. If you ship Claude output in your own product, you cannot verify the mark yourself unless you qualify — you take Anthropic's word that it is there.

The Sonnet 4.6 catch#

Of the three models added on November 3, only Sonnet 4.6 carries a caveat — and it is the detail most likely to bite in production. The release note says that on Sonnet 4.6, text generated for a request that sets temperature, top_p or top_k to a non-default value 'might not carry it.' Anthropic's watermark documentation gives one instruction for anyone who relies on the mark: leave all three settings at their defaults.

That is a real constraint, because tuning sampling is routine in production systems — teams lower temperature for deterministic output or raise it for variety, and frameworks often set these values on the developer's behalf. A product that carefully ships Sonnet 4.6 output under the assumption that it is marked could be shipping unmarked text without knowing it. If the mark matters to your compliance story, audit every temperature, top_p and top_k in the call chain, including the ones your framework adds silently.

Why now: the December 2 deadline#

The legal driver is Article 50(2) of the EU AI Act, which requires providers of generative AI systems to mark their outputs in a machine-readable, detectable way. The rule has applied since August 2, 2026, with a transition period to December 2 for systems already on the market before that date. Anthropic signed the EU's Code of Practice on Transparency of AI-Generated Content in July 2026, one of around 190 signatories, and explained the watermark system in a post published August 14 and updated September 1.

Anthropic is applying the mark worldwide, not just in the EU, because — in its own words — it does not yet have a durable way to scope it by region. The same pattern is spreading across the industry: xAI's developer documentation says text from Grok API models has carried a similar invisible watermark by default since November 7. Other major model developers have signed the same Code of Practice and are implementing their own marks.

What to watch#

Three things to track from here. First, the remaining four 4.x models — Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 — have until December 2 under the transition period, and Anthropic's own support page is the place to watch for their status changing. Second, whether detection access widens: as long as only vetted organizations can check the mark, the public has to take the system on trust, and claims about AI-generated text in the wild stay unverifiable for everyone else. And third, the Sonnet 4.6 sampling caveat — whether Anthropic closes it, or whether developers simply learn to leave the defaults alone.

Anthropic's support page carries one line developers should not miss: if you deploy Claude in your own product, you should independently assess what Article 50 requires of your products and services. The provider's mark covers the model's output. What your product does with that output — and what the law asks of you for it — is your question to answer.