Claude Fable 5.1's watermark skips most code

Anthropic began embedding an invisible watermark in every response from Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, but the technique barely reaches code output. Anthropic says the watermark works by shifting the randomness Claude uses when several next-token choices would produce equally valid text, a method that leaves natural-language prose statistically detectable without changing its meaning.

Code doesn't give the model that same freedom. Swapping a variable name, operator, or function call can change what a program does or break it outright, so Anthropic skips the watermark whenever a token is required for correctness. The company also confirmed the mark is weak on short passages, factual text, and lightly edited drafts, not just code.

The watermark exists to satisfy the EU AI Act's Code of Practice on Transparency of AI-Generated Content, which Anthropic signed in July 2026 and which applies to any model the company released after August 2, 2026. Anthropic is applying it worldwide rather than restricting it to EU users, since it says it has no reliable way yet to scope the feature by region. A detection API is rolling out in private preview, limited for now to regulators, law enforcement, media, researchers, and organizations with their own compliance obligations under the Act.

For developers pasting Claude-generated code into a codebase, the practical result is that the presence or absence of a watermark says little about whether that code came from the model. A generated function can be functionally identical to hand-written code and carry no detectable signature at all.