AI FRONTIER · SIGNAL REPORT

Change One Parameter and Claude's Watermark Is Gone

Anthropic emailed API customers that watermarking is being extended to four models including Mythos 5, but Opus 4.6 and Sonnet 4.6 are watermarked only while sampling parameters stay at defaults, so changing a single parameter bypasses it.

ISSUE 128 · 2026-10-31 · SOURCE / The Decoder report · STATUS / VERIFIED · TOKENTA SIGNAL RESEARCH

ANTHROPIC × WATERMARKREPORTED + ANALYSIS

IMAGE: The Decoder report — FROM THE ORIGINAL REPORT ↗

MODELS ADDED

Claude models covered by this watermark expansion.

SAMPLING DIALS

Sampling parameters that decide whether older models are watermarked; changing one bypasses it.

December

EU DEADLINE

The month of the EU AI Act transparency deadline.

KEY TAKEAWAYS

01THE EMAIL — Anthropic emails API customers announcing the watermark expansion.
02THE FOUR — Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 are brought into coverage.
03THE CAVEAT — Opus 4.6 and Sonnet 4.6 are watermarked only under default sampling parameters.

WHAT HAPPENEDWhat Happened

Anthropic is extending text watermarking to four models: Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Models released after August 2, 2026 ship with the watermark built in; older models are being patched in batches, and these four are part of that effort, against the backdrop of the EU AI Act's December transparency deadline.

The watermark works by weaving an invisible statistical pattern into the randomness of word choice. Anthropic says it adds no characters or tokens, does not change price or request format, and does not alter meaning, quality or readability, and customers cannot switch it off directly.

The loophole sits in the two older models. Opus 4.6 and Sonnet 4.6 are watermarked only when all three sampling parameters - temperature, top_p and top_k - stay at their default values. Change any one of them, even slightly, and the output carries no watermark.

ONE DIALOne Dial Is Enough to Bypass the Compliance Mark

Sampling parameters control how much randomness goes into word choice, and they are things developers adjust every day. Raise the temperature a little and output gets more divergent; lower it and output gets more stable. All of that is normal usage. The result, on the two older models, is that normal tuning takes the watermark away with it.

Anthropic was blunt in the email: customers relying on the watermark for compliance should keep all three parameters at their defaults. That effectively pushes compliance responsibility back onto the caller; the mark itself is no longer an unconditional guarantee.

THE DISCOUNTThe Watermark Promise Is Being Discounted by Reality

The watermarking technique comes from Google DeepMind's SynthID text method, which leaves traces inside randomness. The premise for those traces to work is that the random process behaves as expected. Once a developer changes the sampling parameters, the random process changes, and the statistical basis of the trace changes with it.

This is why new and older models must be read separately. Opus 5.5 and newer models already carry the mark, and the email adds no such restriction for them. The patching of older models looks more like an engineering compromise to meet a deadline: coverage first, unconditional guarantees later.

ASK VERSION FIRSTWhat Compliance Teams Should Really Ask Is Version and Parameters

For any team relying on watermarks for compliance, the checklist now has an extra line. First confirm which model is running, then confirm whether temperature, top_p and top_k are at their defaults; miss either condition, and the mark's presence on the output cannot be assumed.

The steadier approach is to move directly to a newer model, or to lock the three parameters down at the calling layer. The watermark detection API is still in private preview, and after-the-fact detection cannot fill the gap left by output that was never marked.

THE SHIFTWhat Changed

FROM ALWAYS TO IF

From unconditional to parameter-dependent

The watermark status of older models depends on whether the caller tuned the parameters.

FROM MODEL TO CONFIG

From model capability to call configuration

The first item in a compliance check has become version and parameters.

SIGNAL JUDGMENT — ANALYSISA Watermark Is Not a Switch; It Is a Conditional Promise

Tokenta Signal's judgment: a watermark's value lies in being unconditional, and Anthropic's patch this time exposes precisely the condition. Compliance cannot rest on the assumption that nobody tunes parameters; version and parameters must be checked together.

WHAT REMAINS UNCERTAINWhat Remains Uncertain

The boundaries of this report: news of the watermark expansion comes from an email Anthropic sent to API customers, reported by The Decoder; Anthropic's public model table has not yet been updated with these four models or the November 3 date. The email not mentioning the same restriction for Mythos 5 and Opus 4.7 does not prove they are watermarked under every parameter setting. The detection API remains in private preview, and detecting a watermark is not the same as proving a text's origin. This does not constitute investment advice.

EVIDENCEEvidence

The original report confirms Anthropic is extending watermarking to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, with Opus 4.6 and Sonnet 4.6 watermarked only while sampling parameters stay at defaults, so changing a single parameter bypasses it.

ORIGINAL: Anthropic's watermark rollout leaves a loophole on two older Claude models ↗
the-decoder.com · 2026-10-31 · The Decoder report

PRIMARY SOURCESPrimary Sources

The Decoder: Anthropic's watermark rollout leaves a loophole on two older Claude models
https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/
AI ParaBellum: Anthropic Expands Claude Watermark to Older Models, With a Loophole
https://aiparabellum.com/anthropic-claude-watermark-older-models-loophole/

RELATEDRelated Intelligence

CORRECTIONS: if a fact on this page is wrong, we keep a correction record on the page. Contact team@tokenta.space