Anthropic adds a text watermark to four Claude models on 16 November

Anthropic told its API customers (the API is the connection that software uses to operate a model) that Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 get a text watermark from 16 November 2026. Opus 4.6 was due on 3 November. Anthropic gave no cause for the new date.

How it works

The watermark is a pattern in word selection that a tool can find. It is part of each model. It works on each cloud that hosts the models. Customers cannot set it off. Anthropic says it does not change meaning, quality or price.

The parameter limit

The mark applies only if temperature, top_p and top_k stay at default. These three parameters control how random the word selection is. Anthropic tells customers who must have the mark for compliance to keep them at default. Anthropic gave no test results for its quality claims.

What others say

Blogger John Gruber argues that a pattern from a watermark cannot select words as correctly as meaning can. The trade outlet Artificial Lawyer says law firms with clients who ban AI use could find the mark.

Insight: The mark is built for a date, not for strict proof

In the first group on 31 October, 2 of 4 models had the parameter limit. In the new group all 4 have it. Anthropic gives the EU AI Act date in December as the cause, and OpenAI ships its own mark switched off for API customers. This looks like a race to show a mark by the deadline, while the customer must keep it working. Tell your vendor to show who proves the mark.