Anthropic is expanding its text watermark to four more SI models, but API users can bypass the mark on two of them by tweaking a single sampling parameter. The move is part of a broader effort to comply with the EU AI Act's December deadline, which requires transparency in generated content.

What Happened

In an email to API customers, Anthropic announced it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Newer models like Claude Opus 5.5 already carry the mark, while the rollout on Amazon Bedrock may take a few extra days. The company describes the watermark as an "imperceptible pattern in word choice" that does not change meaning, quality, or readability. Anthropic states that the addition adds no extra characters or tokens, does not affect pricing or request formats, and has a negligible impact on output speed. The company also clarifies that the watermark "contains no information about the user, their organization, or their conversations with Claude."

A significant caveat applies to the older models, Claude Opus 4.6 and Claude Sonnet 4.6. Anthropic says the watermark is applied only when developers leave the three sampling parameters—temperature, top_p, and top_k—at their default values. These parameters control the randomness used in word selection. A minor tweak to any one of them is sufficient to dodge the watermark. Anthropic warned in the email, "If you rely on the watermark for your own compliance, keep these settings at their defaults." The email did not mention this caveat for Claude Mythos 5 or Claude Opus 4.7, suggesting the watermark applies to those models regardless of sampling settings. Notably, Opus 4.6 remains Anthropic's best writing model, outperforming even the newer Opus 5.5 in some metrics.

Anthropic uses a variant of Google DeepMind's SynthID text method, which adjusts the randomness behind word selection to create a statistically detectable pattern. The company has a detection API in private preview for verifying the watermark. This rollout aligns with regulatory pressures; every model released after Aug. 2, 2026, shipped with the watermark on day one, while older models are receiving it in stages. Since the watermark is embedded at the model level, it functions across all platforms hosting these SI models, including Amazon Web Services, Google Cloud, and Microsoft Foundry, and customers cannot disable it.

Why It Matters

The loophole on Opus 4.6 and Sonnet 4.6 creates a potential compliance gap for developers who prioritize specific output characteristics over regulatory transparency. By adjusting temperature, top_p, or top_k, users can generate text that appears unwatermarked, potentially bypassing detection systems that rely on the standard pattern. This is particularly relevant for Opus 4.6, which is widely regarded as Anthropic's strongest writing model, making it a preferred choice for high-quality content generation despite the availability of newer versions.

The rollout also highlights the divergent approaches to SI watermarking among major labs. OpenAI launched its own watermarking system, textGrain, in October 2026, but it only activates by default for ChatGPT and Codex users in the EU. API customers worldwide can toggle it on or off, and it ships disabled by default. In contrast, Google DeepMind has open-sourced SynthID for text and integrated it directly into its Gemini models. The debate over watermarking continues, with critics like blogger John Gruber arguing that selecting words based on a watermark key rather than semantic fit inevitably degrades precision, as no two synonyms are exactly identical. Legal analysts also raise concerns for sectors like law firms, where clients who have banned SI use might detect the marks, and multiple watermarks in a single document could create legal complications.

The Bottom Line

Anthropic is retrofitting older SI models with watermarks to meet EU AI Act requirements, but developers using Claude Opus 4.6 and Sonnet 4.6 can easily bypass the mark by altering sampling parameters. While newer models enforce the watermark regardless of settings, the loophole in the older, high-performance models poses a challenge for consistent compliance verification across the SI industry.