Anthropic widens Claude text watermark rollout while Opus 4.6 and Sonnet 4.6 allow bypass via sampling defaults
Anthropic is expanding Claude text watermarking in a second wave that The Decoder says will reach API customers starting November 16, 2026, covering Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 as the company moves ahead of the EU AI Act December deadline. On Opus 4.6 and Sonnet 4.6, reporting tied to customer email says the mark applies only when temperature, top_p, and top_k stay at defaults, so non-default sampling can remove it. Anthropic reportedly responded with a compliance warning urging teams that depend on the watermark to keep those settings rather than eliminating the bypass in product behavior. That limitation matters for operators who treat vendor marks as durable provenance under regulation. The packet is excerpt-based and does not include independent technical verification of how the watermark is embedded or enforced.
Anthropic widens Claude text watermark rollout while Opus 4.6 and Sonnet 4.6 allow bypass via sampling defaults
The Decoder reports Anthropic emailed API customers about a second watermark wave starting November 16, 2026, for Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 ahead of the EU AI Act December deadline. On Opus 4.6 and Sonnet 4.6 the watermark applies only when temperature, top_p, and top_k remain at defaults, and Anthropic warns teams that rely on the mark for compliance to keep those settings.
Key takeaway
A wider Claude watermark rollout ahead of EU timing still leaves an easy API-side off switch on Opus 4.6 and Sonnet 4.6 when sampling leaves defaults.
What happened
The Decoder reports that Anthropic emailed API customers about a second watermark wave beginning November 16, 2026, for Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5, framed ahead of the EU AI Act December deadline.
On Opus 4.6 and Sonnet 4.6, the same reporting says the watermark applies only when temperature, top_p, and top_k remain at defaults; changing a sampling parameter can disable it, and Anthropic warns compliance-reliant teams to keep default settings.
Evidence
Anthropic planned a second watermark wave for four Claude models starting November 16, 2026.
The Decoder · attributed
The Decoder reports Anthropic emailed API customers about a second watermark wave starting November 16, 2026, for Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 ahead of the EU AI Act December deadline.
On Opus 4.6 and Sonnet 4.6 the watermark is conditional on default sampling settings.
The Decoder · attributed
On Opus 4.6 and Sonnet 4.6 the watermark applies only when temperature, top_p, and top_k remain at defaults, and Anthropic warns teams that rely on the mark for compliance to keep those settings.
Excerpt reporting says API users can disable the watermark on older models by changing one sampling parameter.
The Decoder · attributed
But on older models like Claude Opus 4.6 and Sonnet 4.6, API users can disable the watermark by changing a single sampling parameter.
Anthropic's public response to the bypass reporting is characterized as a compliance warning.
The Decoder · attributed
Anthropic's response to the loophole? A compliance warning.
Why it matters
Regulatory and enterprise provenance programs that assume always-on vendor watermarks may fail audits if API callers can turn marks off with routine sampling changes.
Limits and uncertainties
Reporting in the packet is excerpt-based and does not establish independent verification of watermark mechanics or whether newer models close the sampling bypass.
Practical implications
API operators who need watermark evidence for compliance should document and lock temperature, top_p, and top_k defaults on Opus 4.6 and Sonnet 4.6 rather than relying on the mark alone.
What to watch
Whether the November 16, 2026 rollout changes watermark behavior on Opus 4.6 and Sonnet 4.6 beyond the default-sampling condition described in customer email reporting.
