Anthropic is expanding its text watermark to four more Claude models to meet the December deadline of the EU AI Act, though API users can bypass the mark on two older models by tweaking sampling parameters. In an email to API customers, Anthropic announced that the watermark is being applied to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, while newer models like Claude Opus 5.5 already carry the mark. The company describes the watermark as an imperceptible pattern in word choice that embeds a statistically detectable pattern via a variant of Google DeepMind's SynthID text method without affecting pricing or output formats.
On Claude Opus 4.6 and Claude Sonnet 4.6, however, the watermark only activates when developers keep the temperature, top_p, and top_k sampling parameters at their default values, allowing a minor adjustment to dodge the system. Anthropic warns compliance-reliant users to keep these settings at defaults, though the caveat is not mentioned for Claude Mythos 5 or Claude Opus 4.7. The watermark implementation spans multiple hosting platforms including Amazon Web Services, Google Cloud, and Microsoft Foundry, where customers cannot turn it off. Meanwhile, critics continue to debate the impact of text watermarking on semantic precision and document compliance.
Why it matters: This rollout helps Anthropic comply with the EU AI Act while highlighting technical limitations and user workarounds in current AI watermarking implementations.