Claude Text Watermark Reaches Four Older Models, Holds Only at Default Settings

LONDON (BYTEVYTE) Anthropic plans to extend its Claude text watermark to four older models, Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, with a reported start date of November 3, 2026. The mark is being added to help meet EU AI Act requirements, yet it can reportedly be bypassed on Opus 4.6 and Sonnet 4.6 by changing a single sampling parameter. For compliance teams, that makes the Claude text watermark only as reliable as each customer's configuration.

The rollout notice reached customers by email, and the text of that email was shared on X rather than published by Anthropic. According to the message, customers who rely on the mark for compliance should keep temperature, top_p and top_k at their default values. Any deployment that tunes those parameters for output style, length or determinism could therefore produce text without the mark, and the customer may not know it.

Why the default-settings condition matters for Claude text watermark compliance

A watermark that depends on default sampling is a weaker control than a property the model guarantees for every output. The vendor controls the mark, but the customer controls the parameters, so responsibility for compliance shifts to the team that sets the API calls.

The four models differ in how the problem applies. The bypass is described for Opus 4.6 and Sonnet 4.6. Coverage of the customer email names all four, but none of the sources state whether Claude Mythos 5 and Claude Opus 4.7 share the same weakness.

What is documented and what is not

Anthropic's public documentation explains what the watermark is for. It does not include independent test results for the older-model rollout, which rests on the customer email. Compliance officers therefore have no third-party evidence showing how reliably the mark survives at default settings, how often it is lost under other settings, or how it performs across the four models.

The sources also do not settle whether the mark is detectable in practice or how it holds up after text is edited. Without published test data, buyers cannot measure the gap between the stated control and the observed behaviour.

Options for teams that depend on the mark

Teams using the four models for EU-facing output have three practical paths. They can pin temperature, top_p and top_k at defaults across every workload that must carry the mark, which limits tuning for creative or structured tasks. They can separate compliance-bound traffic from workloads that need custom sampling and route each to a different model or pipeline. Or they can treat the watermark as one signal among several, backed by their own logging and review, rather than as the compliance control itself.

Each option carries a cost. Locking defaults constrains output quality for some use cases, while splitting traffic adds routing and monitoring work. Relying on the mark alone leaves the compliance position dependent on an email the vendor has not published in its documentation.

Verdict

Decision-makers should first check which of the four models sit in their production stack and whether any integration sets the three named parameters to non-default values. They should then ask Anthropic for the test basis behind the default-settings condition and watch whether the November 3 start date comes with published documentation. Until independent results or formal documentation appear, the watermark should not be the sole basis for an EU AI Act compliance claim.

While we strive for accuracy and rigor, ByteVyte can make mistakes. Readers are advised to verify all information independently and conduct due diligence before making decisions based on this reporting. We accept no liability for errors, omissions, or decisions made in reliance on this content.

Photo by Brecht Corbeel on Unsplash

Related Articles

• Claude text watermarking goes global as Anthropic meets EU transparency rules
• Anthropic Debuts Claude Opus 4.7 with Enhanced Reasoning and Visual Capabilities
• Claude Alignment Update: How Unsafe Test Models Hit Real Systems

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.