Claude Fable 5.1 and Mythos 5.1 are Anthropic's first models to watermark text outputs; a detection API is available to eligible groups as required under EU law (Ben Patterson/PCWorld)

Anthropic has unveiled its latest Claude models, Claude Fable 5.1 and Claude Mythos 5.1, both with a unique twist - they now include invisible watermarks in text outputs. This move aligns with the EU AI Act's requirements. The Claude Fable 5.1 is available to the public, while Claude Mythos 5.1 is restricted to Anthropic's "trusted access program."

Both models showcase significant performance improvements, particularly in areas like scientific research, coding, and business workflows, with double-digit increases in several benchmarks. Anthropic has also incorporated advanced safety and alignment procedures into these models, developed in response to concerns about models potentially accessing resources outside their designated environments.

The standout feature of these models is the watermarking of text and file outputs. This watermark, undetectable to the human eye, doesn't affect the quality of the output and contains no user or conversation information. The detection API, required for accessing the watermark, is accessible to selected groups under EU law. This includes regulators, law enforcement, media, fact-checkers, independent researchers, and other qualified entities.

The watermarking process works by subtly adjusting the model's word choices during inference, nudging it towards slightly different word options. This method ensures the watermark survives basic text editing, yet remains invisible. The presence of the watermark indicates Claude's involvement, but its absence doesn't necessarily prove human authorship.