Anthropic's text watermark can be skipped on two older Claude models

What Anthropic announced

Anthropic is widening its text watermark to cover four more Claude models: Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. As the-decoder.com reported, the company told API customers about the change by email. Newer models such as Opus 5.5 already carry the mark. On Amazon Bedrock, the cloud service that resells Claude, the rollout may take a few extra days.

A watermark here is not a visible label. Anthropic describes it as an imperceptible pattern in word choice. It says the pattern doesn't alter meaning, quality or readability, adds no extra characters, leaves pricing and data formats untouched, and barely slows the model down.

Anthropic also says the pattern carries nothing about who the user is, which organization they belong to, or what they discussed with Claude. It is a signal about the text, not about the person.

The loophole on Opus 4.6 and Sonnet 4.6

The catch concerns the two older models. According to Anthropic, the watermark switches on only when developers leave three settings at their factory values. Those settings are called temperature, top_p and top_k, and they control how much randomness the model uses when it picks the next word.

Changing even one of them is enough to avoid the mark. Anthropic's email tells customers that if they depend on the watermark for their own compliance, they should leave those settings alone.

The email does not include that warning for Mythos 5 or Opus 4.7. That suggests the mark on those two applies whatever the settings are, though the source frames this as an inference rather than a stated fact.

The gap is notable because Opus 4.6 is still Anthropic's best writing model, ranked above even Opus 5.5. People who care most about prose quality may be the ones most tempted to adjust the randomness dials.

How the watermark works

Anthropic uses a variant of SynthID for text, a method from Google DeepMind. Language models write one word at a time, choosing from several plausible candidates. SynthID nudges that choice in a hidden, structured way, so the finished text carries a statistical fingerprint. A reader sees nothing, but a detector can tell whether the pattern is there.

This also explains why the settings matter. The method works by shaping the model's randomness, so changing the randomness controls can interfere with the pattern. Anthropic has not detailed the mechanics beyond what is reported, so treat that as a plausible reading rather than a confirmed explanation.

Because the mark is built into the model itself, it follows the model to every platform that hosts it, including Amazon Web Services, Google Cloud and Microsoft Foundry. Customers cannot turn it off. Anthropic has a detection tool in private preview for checking text, and more detail sits on its content labeling page.

Why now, and how rivals compare

Anthropic says the EU AI Act's December deadline is behind the push to mark older models. Every model released after Aug. 2, 2026 has shipped with the watermark from the first day, while older ones are being brought in step by step.

Competitors are handling it differently. OpenAI launched its own system, textGrain, in October 2026. It is on by default only for ChatGPT and Codex users in the EU, and API customers anywhere can toggle it. It ships switched off. Google DeepMind has released SynthID for text as open source and built it into Gemini models.

The result is a mixed picture: always-on for some Anthropic models, conditional for others, optional at OpenAI.

Who is affected and what critics say

The most exposed group is businesses that build products on the API and pin them to a specific model version for stability. Our newsletter noted that many enterprise setups do exactly this with Opus 4.6 and Sonnet 4.6. A product could end up with no watermark simply because someone changed a randomness setting long ago.

Not everyone accepts the quality claim. Blogger John Gruber argues that synonyms are never perfectly interchangeable, so selecting words by a hidden key rather than by best fit must cost some precision.

The legal outlet Artificial Lawyer raises a different worry. Law firms whose clients forbid AI use could find that a mark reveals AI involvement. Several watermarks stacked inside one document could add further complications.

• Developers: check which model and settings you use.
• Compliance teams: don't assume the mark is present on the two older models.
• Internal drafting only: exposure is likely smaller, but confirm with whoever owns compliance.