Anthropic will extend its invisible text watermark to four older Claude models beginning November 3, according to an email the company sent to API customers that was first reported by German tech outlet The Decoder. The move brings Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 under the same content-labeling regime that newer models have shipped with since August, and it lands ahead of the EU AI Act's transparency deadline in December.

Which Models Are Getting the Watermark For more context on this story, see our ongoing breaking AI news.

The expansion covers four models that predate Anthropic's day-one watermarking policy: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Every model the company has released since August 2, 2026 already carries the watermark out of the box, with Claude Opus 5.5 among the newer models that included the mark at launch.

Anthropic noted in its email that the rollout on Amazon Bedrock, one of the major cloud platforms hosting Claude models, may take a few extra days. Mashable reported that the expansion formally begins on November 3.

How the Watermark Works

Anthropic describes the watermark as an "imperceptible pattern in word choice" that does not change a text's meaning, quality or readability. Technically, the system is a variant of SynthID Text, the watermarking method Google DeepMind open-sourced last year, which subtly shifts how the model selects among candidate words during generation to produce a pattern that statistical detectors can pick up.

The company says the watermark adds no extra characters or tokens, does not affect pricing or request and response formats, and barely slows output speed. Anthropic also states that the mark "contains no information about the user, their organization, or their conversations with Claude."

Because the watermark is embedded at the model level rather than applied by a specific app, it works on every platform that hosts these models, including Amazon Web Services, Google Cloud and Microsoft Foundry. Customers cannot turn it off.

A Loophole on Two Older Models

The Decoder's report highlighted a notable caveat: on Claude Opus 4.6 and Claude Sonnet 4.6, the watermark only activates when developers leave three sampling parameters — temperature, top_p and top_k — at their default values. These parameters control how much randomness the model uses when picking words, and a minor tweak to any one of them is enough to dodge the watermark.

Anthropic reportedly warned customers directly: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

The email did not attach that caveat to Claude Mythos 5 or Claude Opus 4.7, which suggests the watermark applies to those two models regardless of sampling settings. The gap matters in practice because Opus 4.6 remains a widely used model, particularly among developers who favor it for writing tasks.

Anthropic operates a detection API, currently in private preview, that lets customers verify whether a given text carries the watermark. The company points to its content labeling information page for further technical detail.

Why Anthropic Is Doing This Now

The timing is tied to regulation. Anthropic pointed to the EU AI Act's December deadline as the reason it is retrofitting older models, since the law's transparency obligations require that AI-generated content be detectable as such. The company's approach has been to ship watermarks on all new models from August 2, 2026 onward, while bringing older models into compliance in stages.

The November 3 expansion effectively completes that process for the models most enterprises still use in production. Once it lands, essentially the entire current Claude lineup will produce watermarked text by default.

How Rivals Handle AI Text Labeling

Anthropic is not alone in the space, but its competitors have made different choices. OpenAI launched its own watermarking system, textGrain, in October 2026, yet it only activates by default for ChatGPT and Codex users in the European Union. API customers worldwide can switch it on or off, and it ships disabled — a lighter-touch stance than Anthropic's always-on, can't-be-disabled approach.

Google DeepMind has open-sourced SynthID Text and built it into its Gemini models, making it the technological foundation that Anthropic itself is building on.

The Debate Over Whether Watermarks Degrade Writing

Whether watermarking truly leaves text quality untouched remains disputed. Blogger John Gruber has argued that no two synonyms mean exactly the same thing, so choosing words according to a watermark key rather than purely on semantic fit inevitably degrades precision, even if slightly.

Artificial Lawyer, a legal trade outlet, raised a different concern for law firms: clients who have banned AI use in their engagements could discover the mark in work product, and documents that pass through multiple AI systems could end up with layered watermarks that complicate analysis.

What It Means for Developers and Businesses

For most API customers, the change will be invisible: same pricing, same response formats, same latency profile. The practical consequences concentrate in three areas. Compliance teams in the EU get a built-in mechanism for meeting the AI Act's disclosure requirements. Developers who tune sampling parameters for creative or technical reasons need to understand that doing so on Opus 4.6 and Sonnet 4.6 will strip the watermark, which could matter if their own customers expect compliant output. And anyone publishing Claude-generated text should assume it is statistically identifiable as AI-generated going forward.

The expansion also signals where the industry is heading. Anthropic, OpenAI and Google now all operate production watermarking systems, and the differences between them — always-on versus opt-in, universal versus EU-only — are becoming competitive positioning rather than technical footnotes. As the EU AI Act deadline approaches, expect other providers to make their own staging plans public before December.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →