Anthropic to add text watermarking to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 from Nov 16, with a sampling-parameter loophole

Anthropic has emailed API customers to say a second wave of older Claude models (Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5) will carry a statistical text watermark from 16 November 2026, with Amazon Bedrock possibly a few days later. In this wave the watermark only applies when temperature, top_p and top_k are left at their defaults, so changing any one of them avoids it. Customers cannot switch it off. Opus 4.6 was originally in the first wave (planned 3 November) and moved without a stated reason. The first wave was announced on 31 October. The details come from The Decoder's account of the email, and we have not seen the email itself.

Developers and enterprises using these models via the API need to check whether their sampling settings are non-default, because that decides whether their output is watermarked. If the watermark is meant to support EU AI Act compliance, a loophole this simple weakens it, and compliance teams should not treat the default-settings watermark as a dependable control. Anyone relying on watermark detection to spot Claude text will get false negatives from any output generated with tweaked settings. Wider impact is limited: it is a second-wave extension of a policy already announced, and the effect on output quality is disputed and unmeasured.

Claims that quality is untouched (Anthropic) or that it is damaged (critics) are both unsupported by published tests so far. Calling Opus 4.6 Anthropic's best writing model is opinion, and "easy to bypass" is a reasonable reading of the stated condition but has not been independently tested at scale. The unexplained Opus 4.6 delay invites speculation, but nothing in the coverage shows it means anything.

1. 01Whether Anthropic publishes the policy in official documentation or a changelog before 16 November, and whether the default-settings condition is confirmed or changed.
2. 02Independent detection tests after 16 November: how reliably the watermark is detected at default settings, how quickly it disappears with one altered parameter, and any measured effect on output quality.
3. 03Whether the EU AI Act deadline in December brings guidance on what counts as sufficient marking, and whether Anthropic tightens the older-model watermark in response, or whether Bedrock's date slips.