Anthropic’s Claude watermark rollout has a bypass on two older models

Anthropic is expanding text watermarking across more Claude models, according to an email to API customers reported by The Decoder. The move matters for developers and businesses that plan to rely on AI-content labeling for compliance, because two older Claude models can avoid the watermark when API sampling settings are changed.

The company says the new rollout covers 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 include the watermark, and Anthropic told customers that deployment on Amazon Bedrock may take a few additional days.

What Anthropic says the watermark does

Anthropic describes the system as a subtle pattern in word choice that can be detected statistically without changing the content’s meaning, readability or quality. The company also says the watermark does not add characters or tokens, does not change pricing, and does not affect request or response formats.

According to Anthropic, the watermark does not include information about the user, the customer organization or the conversation with Claude. That distinction is important for teams handling sensitive client work: the mark is meant to indicate AI-generated text, not identify who prompted it.

Anthropic also says it has a detection API in private preview for checking whether the watermark is present. For now, that means teams that need formal verification may have to wait for broader access or use the company’s available content-labeling information.

The bypass developers need to notice

The practical caveat is on Claude Opus 4.6 and Claude Sonnet 4.6. On those two models, Anthropic says the watermark is applied only when three sampling parameters remain at their default values: temperature, top_p and top_k.

Those controls shape how much randomness the model uses when choosing words. The Decoder reports that changing even one of the three is enough to avoid the watermark on those older models. Anthropic’s warning to API customers is straightforward: if they depend on the watermark for compliance, they should keep those settings at their defaults.

That is a real implementation issue for developers. Many production apps tune temperature or other sampling settings to make responses more predictable, more varied or better suited to a specific product experience. If those apps run on Claude Opus 4.6 or Claude Sonnet 4.6, teams should review whether their current defaults are actually Anthropic’s defaults.

For creators, agencies and small businesses using Claude through third-party products, the immediate takeaway is less direct but still relevant: content provenance may depend on the model and integration settings, not just the brand name on the assistant. Teams already tracking Claude usage and cost controls may want to pair this with other admin changes, such as Anthropic’s newer reporting tools for Claude spending covered by Emonarc here.

Why older Claude models are being updated now

Anthropic points to the EU AI Act’s December deadline as the reason for adding watermarks to older models. According to the company, every model released after Aug. 2, 2026 shipped with watermarking from day one, while earlier models are receiving it in stages.

Because Anthropic says the watermark is embedded at the model level, it should apply across platforms that host those models, including Amazon Web Services, Google Cloud and Microsoft Foundry. Customers cannot disable it, according to the company.

The exception is not a user-facing off switch, but the sampling-settings behavior on Claude Opus 4.6 and Claude Sonnet 4.6. The Decoder notes that Anthropic’s email does not describe the same caveat for Claude Mythos 5 or Claude Opus 4.7, suggesting those models apply the watermark regardless of sampling settings.

This makes model choice part of compliance planning. If a business needs the watermark to be consistently present, it should confirm both the Claude model and the API parameters in use, especially if it has standardized on an older model for writing quality or product behavior.

How this fits the wider watermark push

Anthropic is using a variant of Google DeepMind’s SynthID text method, which changes the randomness behind word selection to create a pattern that can be detected later. Google DeepMind has open-sourced SynthID for text and built it into Google Gemini models, according to The Decoder.

OpenAI also introduced a watermarking system called textGrain in October 2026. The Decoder reports that it is enabled by default for ChatGPT and Codex users in the EU, while API customers worldwide can turn it on or off and receive it disabled by default.

There is still debate over whether watermarking leaves writing untouched. The Decoder cites blogger John Gruber’s argument that synonym choices are never perfectly equivalent, and that watermark-driven word selection can reduce precision. It also points to concerns raised by Artificial Lawyer that law firms could face complications if clients ban AI use or if multiple watermarks appear in one document.

For Emonarc readers, the lesson is practical: do not treat watermarking as a blanket guarantee. If your product, agency workflow or client deliverable depends on AI-content labeling, check the exact model, hosting platform, API settings and detection process before making promises to customers.