Anthropic announced a second round of text watermark retrofits starting November 16, 2026, extending the feature to Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5. The move comes ahead of the EU AI Act’s December deadline requiring compliance for legacy models. While newer releases like Claude Opus 5.5 shipped with native watermarks on day one, older models are receiving the updates in stages across all hosting platforms, including Amazon Bedrock, Google Cloud, and Microsoft Foundry.
However, Anthropic confirmed that for Claude Opus 4.6 and Claude Sonnet 4.6, the watermark only applies when temperature, top_p, and top_k remain set to default values. Developers can effectively bypass the statistical pattern by altering any single sampling parameter. The company explicitly warned API customers relying on the mark for regulatory compliance to keep default settings intact.
Anthropic relies on a variant of Google DeepMind’s SynthID technology, which slightly adjusts word selection randomness without adding tokens or affecting performance. While models such as Claude Mythos 5 and Opus 4.7 apply the watermark regardless of parameters, the loophole on Opus 4.6 is notable given its continued popularity for complex text generation.
Why it matters
API developers using Claude Opus 4.6 for writing tasks can avoid text watermarking by simply tweaking temperature or top_p values.
Enterprise teams using Anthropic models to meet EU AI Act compliance must ensure sampling parameters stay at default settings.
Model watermarking remains technically fragile, presenting ongoing challenges for companies managing regulatory requirements across multiple host platforms.
Source: the-decoder.com



