The Pulse
Anthropic’s Claude Watermark Leaves a Sampling Loophole
Anthropic plans to add text watermarks to four older Claude models, but says API users should keep default sampling settings on Opus 4.6 and Sonnet 4.6 if they rely on the mark. The caveat gives developers a way to avoid watermarking those
AI.info Team · Written with AI · fact-checked ·
“If you rely on the watermark for your own compliance, keep these settings at their defaults.”
Anthropic, in an email to API customers, as quoted by The Decoder
Anthropic is adding text watermarks to four older Claude models, but its instructions for two of them leave developers a way to avoid the mark. On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark applies only when three sampling settings remain at their default values, The Decoder reported on Oct. 31, citing an email to API customers. Changing any one of those settings can bypass the watermark, according to the report.
Four older Claude models join the rollout
The planned additions are Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Newer models, including Claude Opus 5.5, already carry the watermark, while the rollout through Amazon Bedrock may take a few additional days, Anthropic told customers in the email. Anthropic’s Help Center says the company is adding marks to models released before Aug. 2, 2026.
The settings named in the email—temperature, top_p and top_k—shape how much randomness a model uses when choosing words. Developers using Opus 4.6 or Sonnet 4.6 through the API can alter one of them and, according to The Decoder’s account of Anthropic’s guidance, avoid the watermark. The email did not describe the same default-setting condition for Mythos 5 or Opus 4.7, so the report does not establish whether those models behave the same way when sampling settings change.
A statistical mark, not a visible label
Anthropic describes its text watermark as an imperceptible pattern in a model’s word choices. Its explanation of the method says the system uses a key and preceding words to shape random choices among plausible next words, creating a pattern that can be checked later. The company says the process adds no characters or tokens and does not change meaning, readability or price.
The mark is intended as evidence that Claude likely helped produce text, not proof of who wrote it or how. Anthropic says detection works poorly on short passages and can become unreliable after substantial editing, paraphrasing or translation. Its detector remains in private preview for eligible organizations and enterprises, according to the company’s help page.
Compliance depends on the API configuration
Anthropic says it is introducing the marks in response to commitments under the EU AI Act’s transparency rules. Newer Claude models launched in the EU on or after Aug. 2, 2026, support marking at launch; models released earlier are being brought into the system in stages. The company says marking operates at the model level across supported Claude products and cloud platforms.
That broad deployment promise does not eliminate the practical distinction in the API guidance: customers relying on watermarking for compliance must leave all three sampling controls at their defaults on Opus 4.6 and Sonnet 4.6. The email’s warning gives developers a clear instruction, but it also means the mark’s presence on those two models depends on how a request is configured. Anthropic’s customer guidance, as reported by The Decoder, leaves that condition attached to the API settings themselves.