Anthropic is adding a text watermark to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. The Decoder reports that Anthropic told API customers about the rollout in an email. Claude Opus 5.5 and other newer models already have the watermark; Amazon Bedrock may receive the change a few days later than other platforms.
Anthropic describes the mark as an “imperceptible pattern in word choice.” The company says it does not change meaning, quality or readability, add characters or tokens, affect pricing, or include information about users or their conversations. Anthropic also says output speed is barely affected.
How the loophole works
On Opus 4.6 and Sonnet 4.6, the watermark applies only when developers leave three sampling settings at their defaults: temperature, top_p and top_k. These settings affect how much randomness the model uses when choosing words. Anthropic says changing any one of them is enough to avoid the watermark, and warns customers who rely on it for compliance to keep all three at their default values.
Anthropic’s email does not mention the same limitation for Mythos 5 or Opus 4.7, which suggests the watermark may apply to those models regardless of settings. The email does not explicitly confirm that. The Decoder also describes Opus 4.6 as Anthropic’s strongest writing model, ahead of Opus 5.5.
A compliance tool with limits
Anthropic says it is retrofitting older models ahead of the EU AI Act’s December deadline. Models released after Aug. 2, 2026, shipped with the watermark from launch; older models are being updated in stages. Because the mark is embedded at the model level, Anthropic says it works across hosting services such as Amazon Web Services, Google Cloud and Microsoft Foundry, and customers cannot switch it off directly.
Anthropic has a detection API in private preview. The system is a variant of Google DeepMind’s SynthID text method, which alters word-selection randomness to create a pattern that can be detected statistically. The effect on writing quality is disputed: blogger John Gruber argues that choosing words partly to create a watermark can reduce precision, while legal trade outlet Artificial Lawyer has raised concerns about watermarks complicating documents for law firms.
Why it matters
Developers using Opus 4.6 or Sonnet 4.6 as evidence of AI-generated text for compliance cannot assume the watermark will be present if they change even one sampling setting. Anthropic says the mark does not affect readability, but its effect on writing precision remains disputed.
