Anthropic adds a text watermark to four more Claude models
On two of the four models, a change to one parameter turns the watermark off.
Anthropic told API customers in an email that it will add a watermark to the text of four Claude models, among them Claude Mythos 5. The watermark is a pattern in the selection of words. Anthropic gives the EU AI Act date in December as the cause. Newer models, such as Claude Opus 5.5, have it at this time.
How it works
Anthropic says the pattern does not change the meaning, the quality or the price. It tells that the pattern has no data about the user. The method is a variant of SynthID text from Google DeepMind. Customers cannot turn it off.
The limits
On Claude Opus 4.6 and Claude Sonnet 4.6, the watermark works only if the three sampling parameters stay at default. These control how random the selection of words is. One small change to one parameter stops the mark. Anthropic gave no test results for its quality claims.
What is disputed
Blogger John Gruber says that no two synonyms have the same meaning. The blogger says that a watermark must decrease precision. The outlet Artificial Lawyer says that law firms can have problems if clients forbid AI use and find the mark.
Insight: Anthropic marks all text, OpenAI marks mostly in the EU
OpenAI started its textGrain watermark in October 2026. It is on by default only for ChatGPT and Codex users in the EU. API users can set it on or off. Anthropic marks at model level, on all clouds. This reads as two ways to meet one EU law: Anthropic uses one setup for all clouds, and OpenAI keeps the alternative. If you work in the EU, find which mark each vendor gives you.
Sources
Posted
