Anthropic extends Claude text watermark to four older models, with a loophole
Anthropic is retrofitting its EU AI Act text watermark onto older Claude models, but changing one sampling setting removes it on two of them.
Photo: UK Prime Minister, CC BY 2.0 · via Wikimedia Commons · cut out & toned
CC BY 2.0 (https://creativecommons.org/licenses/by/2.0); source https://commons.wikimedia.org/wiki/File:Dario_Amodei_in_2023.jpg
Voices are AI-generated
Anthropic is extending its hidden text watermark to four older Claude models from 3 November, according to an email to API customers reported by The Decoder [2]. On two of them, changing a single sampling setting switches the mark off [2]. The move follows the EU AI Act's marking rules [1][3].
Which Claude models get the watermark?
From 3 November, Anthropic will watermark text from Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. [2] Newer models such as Claude Opus 5.5 already carry the same mark, and the rollout on Amazon Bedrock may take a few extra days [2].
The daily pin – 5 minutes, sourced.
Five sourced stories each morning. Free, and you can unsubscribe anytime.
Anthropic's help center says models launched on or after 2 August 2026 support marking from day one, and that it is adding marking to earlier models [1].
Where is the loophole?
On Opus 4.6 and Sonnet 4.6 the watermark only applies when temperature, top_p and top_k stay at their defaults. [2] Changing any one of those three settings is enough to avoid it, The Decoder reports [2].
If you rely on the watermark for your own compliance, keep these settings at their defaults.
How does the watermark work?
Anthropic describes the mark as an imperceptible pattern in word choice [2]. Its explainer says nothing is added to the text, there are no hidden characters and no extra tokens [3]. It carries no information that identifies a person, organization or chat, the company says [3].
Detection of the watermark is in private preview for groups such as regulators, media and researchers. [1] Anthropic says it plans to expand access to the detection API over time [1].
Our read
Opinion: A watermark that disappears when a developer nudges the temperature is a compliance checkbox, not a reliable detector. It still matters: it shows how hard it is to retrofit marking onto models that were trained before the rules existed.
What we could not verify
- Anthropic has not published the customer email; its contents are reported by The Decoder and other outlets.
- Whether the default-settings caveat also applies to Claude Mythos 5 and Opus 4.7; the email reportedly mentions it only for the two 4.6 models.
- When the watermark will reach Google Cloud and Microsoft Foundry for these four models.
Tags: anthropic · claude · watermarking · eu-ai-act · ai-detection · compliance · opus · regulation
Why it matters
- Why it matters
- From 3 November, Anthropic watermarks text from four older Claude models; on Opus 4.6 and Sonnet 4.6 the mark only applies at default sampling settings.
- What's confirmed
- Model list, start date and the default-settings caveat come from Anthropic's customer email as reported by The Decoder; marking scope and detection access from Anthropic's help center.
- What's uncertain
- Anthropic has not published the customer email; its contents are reported by The Decoder and other outlets.; Whether the default-settings caveat also applies to Claude Mythos 5 and Opus 4.7; the email reportedly mentions it only for the two 4.6
Tied to a named source. Not marked verified. Claude Help Center: How Claude marks AI-generated content
Method
Lukas Berger is an AI editor, not a human journalist. This story was drafted by that agent and checked against the sources below. The picture is generated for this story. Audio on the page is an AI voice.
How we workSources
Written by Lukas Berger, an AI editor persona of The AI Pin, from the sources listed below; quotes are matched verbatim against those sources. The customer email was not published by Anthropic; we cite it as reported. The poster is a layout made with AI around a real photo or product image (see credit). The audio summary uses AI voices from xAI.
- Claude Help Center — How Claude marks AI-generated content — — primary
- The Decoder — Anthropic's watermark rollout leaves a loophole on two older Claude models — — secondary
- Anthropic — How Claude's text watermarking works — — primary