AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic's watermark rollout leaves a loophole on two older Claude models

Collected Oct 31, 2026

Anthropic is adding its text watermark to four more Claude models, according to an email sent to API customers. The models covered are 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 carry the mark. Anthropic says the rollout, including on Amazon Bedrock, may take a few extra days on some platforms.

The company describes the watermark as an imperceptible pattern in word choice that leaves meaning, quality and readability unchanged. According to Anthropic, it adds no extra characters or tokens, does not affect pricing or request and response formats, and only marginally slows output. Anthropic also states that the mark contains no information about the user, their organization or their conversations with Claude. Because it is embedded at the model level, it applies on every platform hosting these models, including Amazon Web Services, Google Cloud and Microsoft Foundry, and customers cannot switch it off.

The most consequential detail is a caveat attached to the two oldest models. On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark is applied only when three sampling parameters — temperature, top_p and top_k — are left at their defaults. Those parameters govern how much randomness the model uses when choosing words. Adjusting any one of them is enough to avoid the watermark. Anthropic's email warns customers that if they rely on the watermark for their own compliance, they should keep those settings at their defaults. The same caveat is not mentioned for Mythos 5 or Opus 4.7, which suggests — though the email does not say so directly — that the mark holds on those models regardless of sampling configuration.

That gap matters because of where Opus 4.6 sits in Anthropic's lineup. The article notes that Opus 4.6 remains Anthropic's strongest model for writing, ahead of Opus 5.5. In practice, developers who want the writing quality of that model and also want fine control over randomness face a genuine trade-off: tuning sampling for creativity, determinism or reproducibility could silently drop the mark. This is my own reading, but it looks like an artifact of retrofitting older models rather than a deliberate design; a watermark baked in at training or serving time on newer models behaves differently from one that depends on default inference settings.

For verification, Anthropic has a detection API in private preview and maintains an information page on content labeling. Detection is the other half of any watermarking scheme: a mark is only useful to a compliance reviewer if there is a reliable way to confirm it. Developers who plan to rely on the signal should expect to build against that API once it is generally available.

The timing is driven by regulation. Anthropic points to the EU AI Act's December deadline as the reason it is watermarking older models now. Models released after Aug. 2, 2026 shipped with the watermark enabled from the start, while earlier models are being brought into line in stages. That staged approach is why the caveats differ across the fleet.

Anthropic's implementation is a variant of Google DeepMind's SynthID text method, which biases the randomness behind word selection to produce a pattern that statistical tests can detect. OpenAI launched its own system, textGrain, in October 2026, but it activates by default only for ChatGPT and Codex users in the EU; API customers worldwide can enable or disable it, and it ships disabled. Google open-sourced SynthID for text and built it into the Gemini models, which makes Anthropic's approach the most centrally enforced of the three for API traffic — apart from the older-model exception.

Whether watermarking truly leaves text untouched is contested. Blogger John Gruber has argued that no two synonyms carry exactly the same meaning, so selecting words by a watermark key rather than by semantic fit inevitably dulls precision. Artificial Lawyer raises a separate concern for law firms: clients who have banned AI use might discover the mark, and several watermarks layered inside a single document could create complications. Both points are arguments about consequences rather than demonstrated failures, but they are worth weighing for anyone in a regulated or high-trust setting.

Why it matters: Teams calling Claude through the API need to know that compliance-by-watermark is conditional on model and settings. If you depend on the mark, pin Opus 4.6 and Sonnet 4.6 to default sampling parameters or move to a newer model where the caveat does not appear, and plan to integrate Anthropic's detection API when it becomes available. Legal and regulated users should treat watermark presence as one input among several, not a guarantee, especially where documents mix outputs from multiple models or vendors.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Anthropic is expanding its text watermark to four more Claude models to meet the EU AI Act's December deadline. On the older models Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by tweaking a single sampling parameter. Anthropic addresses the loophole with just a compliance warning. The article Anthropic's watermark rollout leaves a loophole on two older Claude models appeared first on The Decoder .