On October 31, 2026, THE DECODER reported that an email Anthropic sent to API customers described a text-watermark exception for Claude Opus 4.6 and Claude Sonnet 4.6: changing any one of temperature, top_p, or top_k from its default could avoid the mark. The report’s account makes the exception specific to those two models.
The reported exception names three sampling settings
THE DECODER said the email named Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 for watermarking. Its account tied the settings exception to Opus 4.6 and Sonnet 4.6. It also attributed advice to Anthropic that API customers who rely on watermarking for compliance should keep all three settings at their defaults.
The reported exception adds a model-specific caveat to NeoTeo’s earlier coverage of Anthropic’s Claude watermark plans.
What a watermark result can tell you
Anthropic describes its text watermark as an imperceptible pattern in word choice. A detected mark may mean Claude generated or processed text, such as by proofreading or translating it; the mark alone does not establish that Claude originated the whole passage. Anthropic says its detector is in private preview for eligible organizations and enterprises with relevant compliance obligations. Anthropic’s help page explains the watermark and its detection limits.
A negative result also cannot establish human authorship. Anthropic says detection can fail when marking is unsupported, text is heavily edited, paraphrased, translated or combined with other writing, or the passage is very short. The watermark travels with copied text and may persist through some editing, but those limits mean a detector result is not a complete record of who wrote or processed every part.
Anthropic’s target for older-model marking
Anthropic says marking for older models is in progress and sets December 2, 2026, as its target for covering the models included in that rollout. The date is a target, while the reported exception concerns settings on Opus 4.6 and Sonnet 4.6.
