--- title: "Anthropic Watermarks 4 Older Claude Models, With a Loophole" description: "Anthropic is adding its hidden text watermark to Claude Opus 4.6, Sonnet 4.6, Opus 4.7 and Mythos 5, but one sampling tweak skips it on two of them." url: https://publishmyblog.com/ai/anthropic-claude-watermark-rollout/ section: "AI" author: "Saeed Shahriar" published: 2026-11-02T13:05:00.000Z --- # Anthropic Watermarks 4 Older Claude Models, With a Loophole If you publish text drafted with Claude, your older models are about to mark it too. Anthropic is adding its invisible text watermark to four older Claude models. But on two of them, Claude Opus 4.6 and Sonnet 4.6, **one changed sampling setting skips the mark**. Publishing your posts or client copy drafted with Claude? Then this reaches your text too. ## What Anthropic told customers [The Decoder reports](https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/) that Anthropic emailed API customers about **Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6**, all four released before the August 2, 2026 cutoff. Newer models such as Opus 5.5 already carry the same mark, so for them nothing changes this week. The start date comes from [Pasquale Pillitteri's write-up of the same email](https://pasqualepillitteri.it/en/news/27785/claude-watermark-november-3-opus-4-7-sonnet-4-6): **November 3, 2026, on the Claude Platform**. Amazon Bedrock may take a few more days. It won't cost you more. Anthropic says the mark adds no characters or tokens to your output and doesn't change your pricing. **You can't switch it off**, The Decoder says, and because it sits in the model itself it follows Claude onto every platform that hosts these models. ## The gap on two models On Opus 4.6 and Sonnet 4.6, the mark only applies while three sampling settings stay at their defaults: - `temperature` - `top_p` - `top_k` **Change any one of them and your output goes unmarked.** Check your API calls. Anthropic's email tells anyone who relies on the mark for compliance to "keep these settings at their defaults." The email gives no such caveat for Mythos 5 or Opus 4.7. My read: Opus 4.7 [already accepts only the default sampling values](https://publishmyblog.com/ai/claude-api-drops-sampling-controls/), so **there's nothing left for you to tweak** there. ## How the mark and detection work [Anthropic's explainer on its text watermark](https://www.anthropic.com/news/claude-text-watermark) calls it a version of Google DeepMind's SynthID-Text. **A secret key picks between equally good words** instead of plain chance. Detection needs that key, and Anthropic says the mark is weak in a few places: - short passages - light proofreading of your own draft - code, where Claude has few free choices Still, **a translation by Claude carries the mark**, because Claude chose every word of it. Who can check your text today? Only organisations in the **private preview of the detection API**. [Anthropic's help page](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) lists who qualifies: - regulators, law enforcement and media - fact-checkers and independent researchers - educational organisations and EU civil society groups - enterprises with their own compliance duties Images and other supported files, such as PNG and JPEG, get signed C2PA metadata instead, which anyone can check with Anthropic's Claude Content Checker. Why now? **Anthropic points to the EU AI Act's December deadline**, The Decoder says. Models released after August 2, 2026 had it from day one. ## Why it matters If you write or edit with Claude, picking an older model won't keep your text clean, because **from November 3, 2026 those models mark it too**. So who sees it? Anthropic says the mark carries nothing about you or your chats, but **a client's compliance team in the private preview** can test a long post. That's your agency risk. Say you deliver 2,000 words to a client who banned AI drafts, their team runs your post through the detector, and the mark you never mentioned shows up in a meeting you weren't in. So I'd do three things this week: - Put your AI use in the contract. - Skip the temperature gap. Anthropic's newer API version already rejects those settings. - Edit your drafts hard. Though keep one limit in mind. A detected mark means Claude was likely involved. **A missing mark proves nothing.** ## Common Questions **Can I turn off Claude's text watermark in the API?** No. You can't switch it off, because Anthropic applies it at the model level, so it follows Claude onto AWS, Google Cloud and Microsoft Foundry too. The one gap is Claude Opus 4.6 and Sonnet 4.6, where changing temperature, top_p or top_k from the default leaves your output unmarked. **Which Claude models start watermarking text on November 3, 2026?** Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 start on the Claude Platform on November 3, 2026, and Amazon Bedrock may take a few more days. Every model released after August 2, 2026 shipped with the mark on day one. **Can a client check whether my article was written with Claude?** Only if they're in the private preview of the detection API, which covers regulators, media, researchers, schools and enterprises with their own compliance duties. Anthropic says a detected mark means Claude was likely involved in the text. ## Sources - [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/), The Decoder - [How Claude's text watermarking works](https://www.anthropic.com/news/claude-text-watermark), Anthropic - [How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content), Claude Help Center - [Claude Text Watermark Arrives on Opus 4.7, 4.6 and Sonnet 4.6 on November 3](https://pasqualepillitteri.it/en/news/27785/claude-watermark-november-3-opus-4-7-sonnet-4-6), Pasquale Pillitteri