Anthropic has notified API customers that it is expanding its covert text watermarking system to four additional models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. The initiative targets compliance with the EU AI Act’s upcoming December deadline, expanding on watermarking that shipped by default on all models launched after August 2, 2026.

However, Anthropic acknowledged a technical limitation regarding older models. On Claude Opus 4.6 and Sonnet 4.6, the watermark only embeds if developers retain default sampling parameters for temperature, top_p, and top_k; modifying any of these parameters dodges the watermark entirely. Newer releases like Claude Mythos 5 and Opus 4.7 reportedly retain the watermark regardless of parameter tweaks.

Using a variant of Google DeepMind’s SynthID technique, Anthropic embeds an imperceptible statistical pattern into word selection without adding extra tokens or modifying API payload formats. While cloud platforms like AWS, Google Cloud, and Microsoft Foundry will enforce the marks at the model level, critics and legal observers warn that watermarking may subtlety degrade writing precision and complicate enterprise compliance.

Why it matters

  • API developers on older Claude models can bypass mandatory watermarking by simply tweaking standard sampling parameters.

  • Enterprise users face potential legal and quality complications as AI providers enforce model-level output tracking for EU compliance.

  • Highlights technical divergence between frontier watermarking enforcement and backward-compatibility on legacy API models.

Source: the-decoder.com