Anthropic extends Claude’s watermarking, with a flaw on two older models
Anthropic deploys a textual watermark on four additional Claude models to meet European AI Act requirements. But on Opus 4.6 and Sonnet 4.6, a simple sampling adjustment still bypasses it.
Anthropic extends its textual watermarking to four additional Claude models, including Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. The company's goal is to prepare its models for the constraints of the European AI regulation, with deployment already effective on recent versions and spread over a few days for Amazon Bedrock. However, a loophole persists on two older models: on Opus 4.6 and Sonnet 4.6, simply changing a sampling parameter can remove the watermark.
A discreet watermark designed for compliance
Anthropic describes this watermark as an “imperceptible pattern in word choice”, meaning a statistical signature embedded in word selection. The company states that it does not alter the meaning, readability, or format of responses, nor does it add any extra characters or tokens. In other words, the user sees nothing unless they specifically look under the hood.
The company also specifies that this marking contains no information about the user, their organization, or their conversations with Claude. It operates at the model level, thus across all platforms hosting these versions, including Amazon Web Services, Google Cloud, and Microsoft Foundry. It cannot be disabled client-side.
The bypass involves three well-known developer settings
The important nuance concerns Claude Opus 4.6 and Claude Sonnet 4.6. According to the email sent to API clients, the watermark is only active on these two models if the temperature, top_p, and top_k parameters remain at their default values. These settings control the randomness in generation: changing just one of them is enough to avoid the watermark. Anthropic therefore warns clients who intend to rely on it for compliance to keep these parameters unchanged.
This limitation does not seem to be mentioned for Claude Mythos 5 and Claude Opus 4.7, implying that the watermark applies regardless of settings. [to be verified]
Why now? The European timeline pushes publishers
Anthropic links this development to the December deadline of the European AI Act. Models launched after August 2, 2026, already include the watermark from release, while older ones are updated in stages. The logic is simple: to make AI outputs traceable without burdening developers’ usage, at least in theory.
The mechanism uses a variant of SynthID, Google DeepMind’s method for text. It acts on the probability of words chosen by the model to produce a statistically detectable pattern. OpenAI launched textGrain in October 2026, but this system is enabled by default only for ChatGPT and Codex within the European Union; API clients can enable or disable it. Google has already integrated SynthID into Gemini and made it available for text.
What companies will need to monitor
The debate is not only about compliance. Some observers question the idea that a watermark is neutral regarding text quality, especially when multiple watermarking systems overlap in the same document. Legally, the issue becomes even more sensitive: a law firm may want to prove that a text was not produced by AI, or conversely preserve a clean production chain. The watermark helps, but it does not eliminate ambiguous uses. And it certainly does not settle all disagreements.
In this context, Anthropic chooses a pragmatic path: making watermarking automatic on the newest models, while retrospectively aligning the older ones. However, the bypass on Opus 4.6 and Sonnet 4.6 shows that a compliance mechanism still sometimes depends on a configuration detail that many developers already change out of habit.
Key points
- Four Claude models receive textual watermarking.
- Opus 4.6 and Sonnet 4.6 remain bypassable.
- The marking targets the AI Act timeline.
- Anthropic speaks of an “imperceptible pattern in word choice.”
- Deployment also extends to Amazon Bedrock.
By the numbers
- 4 models — affected by the extension announced by Anthropic.
- 2 models — Opus 4.6 and Sonnet 4.6, bypassable via settings.
- August 2, 2026 — date after which new models already come watermarked.
- December 2026 — deadline mentioned for the European AI Act.