Clarity for AI buildersSponsored by AI Routing
availability

Anthropic expands Claude text watermarking, with sampling-setting limits on older models

Anthropic expands Claude text watermarking, with sampling-setting limits on older models
Anthropic has told API customers that text watermarking will reach four older Claude models starting November 16, 2026, but only when their sampling parameters remain at default settings, according to The Decoder. The announcement also delays the previously scheduled rollout for Claude Opus 4.6. In a November 4 update, The Decoder reports that the new rollout covers Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5. Deployment on Amazon Bedrock may take a few additional days. Anthropic explicitly confirmed that Opus 4.6 would move from its original November 3 date to November 16, without giving a reason. For all four models in this wave, the watermark applies only when temperature, top_p, and top_k are left at their defaults. These parameters govern randomness and word selection during generation. Changing any one of them from its default therefore avoids the watermark under the conditions Anthropic described. That limits its usefulness as a consistent signal of AI-generated text on these models. The earlier announcement covered Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Only Opus 4.6 and Sonnet 4.6 carried the default-settings restriction in that announcement. Mythos 5, Opus 4.7, and Sonnet 4.6 were absent from the subsequent email; The Decoder inferred that their November 3 rollout likely proceeded as planned, but the supplied report does not independently confirm that deployment. Anthropic describes its watermark as an “imperceptible pattern in word choice.” The company says it does not change meaning, quality, or readability, adds no characters or tokens, leaves pricing and request and response formats unchanged, and has little effect on output speed. It also says the watermark “contains no information about the user, their organization, or their conversations with Claude.” These are company claims rather than independent test results in the supplied evidence. The system uses a variant of Google DeepMind’s SynthID text method, adjusting the randomness behind word selection to create a statistically detectable pattern. Because the watermark operates at the model level, Anthropic says it works across hosting platforms, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers cannot directly switch it off, although the sampling-setting restriction provides a way to avoid it on the affected models. For customers relying on detection as part of compliance, Anthropic’s advice is explicit: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” The company has a watermark-detection API in private preview. Anthropic links the retrofits to what the report describes as the EU AI Act’s December deadline. According to the report, every model released after August 2, 2026, shipped with watermarking from launch, while older models are receiving it in stages. Claude Opus 5.5 is among the newer models already carrying the mark. The report also notes competing approaches. OpenAI’s textGrain system, launched in October 2026, activates by default for ChatGPT and Codex users in the EU, while API customers worldwide can choose whether to enable it, with the feature disabled by default. Google DeepMind has open-sourced SynthID for text and incorporated it into Gemini models. Whether text watermarking leaves writing quality unchanged remains disputed. Blogger John Gruber argues that choosing words partly according to a watermark key can reduce semantic precision. Legal trade publication Artificial Lawyer raises concerns about clients detecting AI use they have prohibited, as well as complications from multiple watermarks within a document. The supplied report includes no independent quality assessment or detection testing to resolve those concerns.