Anthropic will retrofit text watermarking to four Claude models on November 16, 2026, including Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5. On Amazon Bedrock the rollout may take several additional days. The mark works only when temperature, top_p and top_k remain at defaults, so a small change removes it, and that matters for compliance processes built around detection.
Second retrofit wave from November 16
Anthropic notified API customers by email about the second stage of retrofitting older models. Claude Opus 4.6 was originally scheduled for November 3 as part of the first wave with Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6. The company confirmed the shift of Opus 4.6 to November 16 without giving a reason. The other three first-wave models were absent from the new message, which points to rollout on November 3 as planned.
Opus 4.5, Sonnet 4.5 and Haiku 4.5 were not included in the previous announcement and now join Opus 4.6. All four models in the November 16 group carry the same sampling restriction. Watermarking applies only when temperature, top_p and top_k stay at default values. In the first wave the same restriction covered two of four models, Opus 4.6 and Sonnet 4.6, while Mythos 5 and Opus 4.7 appeared to be marked regardless of settings.
Anthropic describes the signal as an imperceptible pattern in word choice that does not alter meaning, quality or readability. The method adds no extra characters or tokens, does not change pricing or request and response formats, and has little effect on output speed. The watermark carries no information about the user, organization or conversation content. It is embedded at model level, operates across Amazon Web Services, Google Cloud and Microsoft Foundry, and customers cannot disable it.
What watermark limits mean for business
For companies that publish client-facing text, the November 16 date creates a clear checkpoint for documentation, contracts and review procedures. Teams using Opus 4.6, Opus 4.5, Sonnet 4.5 or Haiku 4.5 through API or cloud platforms need to record default sampling values in their generation pipelines. Small firms with one product copy flow can lock the three parameters once, while large organizations with many teams will need a shared configuration standard to keep detection consistent.
The bypass path deserves attention during vendor and tool selection. Anthropic warns customers that rely on the mark for compliance to keep the three settings at defaults. A detection API exists in private preview for verification, with further detail on a content labeling page. Until access widens, buyers should ask providers how verification is performed, who holds detection rights, and how output created with altered sampling will be treated in audits.
The marker to watch is enforcement of the EU AI Act deadline in December, which Anthropic cites as the reason for retrofitting models released before August 2, 2026. Later models shipped with the mark from release, while OpenAI launched textGrain in October 2026 and Google DeepMind built SynthID into Gemini models. If regulators accept model-level marks with sampling exceptions, such conditional labeling may become the operating baseline for enterprise AI content.
