Anthropic has extended its invisible text watermark to Claude models released before August 2026: Opus 4.7, Sonnet 4.6 and Mythos 5. According to a customer email reported by several outlets, the rollout began on November 3. It matters to anyone who writes website copy, product pages or emails with Claude.
What we know
- Models with a text watermark now: Anthropic's help center lists Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5, 5, 4.8 and 4.7, Sonnet 5.5, 5 and 4.6, and Haiku 5.5.
- Models without one: the same table shows no text watermark for Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. Some outlets listed Opus 4.6 among the November 3 additions; the official page, checked on November 5, does not.
- Timeline: Anthropic announced the watermark on August 14, 2026 for new models, to meet the EU AI Act transparency rules that apply from August 2. The law gives older systems a transition period, and the company said it would add them "over the coming months".
- Cloud platforms: the watermark is built into the model, so it also applies on AWS, Google Cloud and Microsoft Foundry. The help page says the Amazon Bedrock rollout for Mythos 5, Opus 4.7 and Sonnet 4.6 will be complete by November 10.
- One technical caveat: Sonnet 4.6 output may not reliably carry the watermark when `temperature`, `top_p` or `top_k` is set to a non-default value.
- How it works: Anthropic says it is a version of SynthID-Text. No hidden characters or extra tokens are added; a secret key guides the choice between equally good words. The company reports no effect on quality, speed or cost.
- Who can detect it: only Anthropic's key reads it. Text detection is in private preview for regulators, media, fact-checkers, researchers, educators and companies that must verify it. Neither Anthropic page offers a way to turn the watermark off.
- Files: PNG and JPEG images generated by Claude carry a C2PA Content Credential, which anyone can check with the Claude Content Checker.
What changes and what doesn't
- What changes: text from an older Claude model that used to come out unmarked now carries the watermark. That covers many integrations that pinned a model months ago.
- What changes in practice is small: outside approved organizations, nobody can read the watermark today.
- What doesn't change: ownership or authorship. Anthropic says the watermark does not identify the user or account and does not alter rights under its terms.
- What doesn't change for search: Google's guidance on AI-generated content, published on Search Central in February 2023, focuses on whether content is helpful, not on how it was produced.
- It is not proof either way. Anthropic warns that a watermark only signals Claude was involved, perhaps just to proofread or translate, and that its absence does not prove a text is human-written.
How to tell if it affects you
Check which model you use. In the Claude app it's in the model picker. In a WordPress plugin, a chatbot or a script that calls the API, the model name sits in the settings (for example `claude-sonnet-4-6`). If it's on the watermarked list, your new text already carries it.
Don't pay for "humanizers". Rewriting text with another tool to strip the watermark won't help your rankings. What helps is what no model adds on its own: your business data, real examples, current prices and timelines, and a review by someone who knows the subject.
Short and technical text barely carries it. Anthropic says detection is unreliable on short passages and the watermark is sparse in code, facts and exact answers. A title tag or a meta description won't "give you away".
If you publish for an EU audience, the AI Act also places duties on those who publish AI-generated text on matters of public interest. Check with your adviser whether they apply to you. Keep a simple log: which pages were written with AI, with which model, and who reviewed them.
This article is not legal advice.
Related: ChatGPT to Add an Invisible Text Watermark in the EU
Sources
- Anthropic Help Center, "How Claude marks AI-generated content", checked November 5, 2026
- Anthropic, Claude text watermark announcement, August 14, 2026
- Google Search Central, "Google Search's guidance about AI-generated content", February 2023
- Mashable, via 1023jack.com, "Anthropic Expands Text Watermarking To Older Claude Versions Beginning November 3"
Updates: if Anthropic adds the watermark to Opus 4.6 or other models, or opens text detection to the public, it will be noted here with a link.