Anthropic plans to extend its text watermark to four more Claude versions from 3 November, including Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The change brings more of Claude’s output under a system Anthropic says marks AI-generated text without changing how it reads.
Anthropic Watch analysis
What happened
The expansion applies the watermarking technology already in use in Opus 5.5. Mashable reports that Anthropic describes it as an imperceptible pattern in word choice, with no identifying information about the user or organisation included.
The rollout is due to begin on 3 November. The report says the initial deployment covered only the latest Claude versions; the four named models are now being added.
Why it matters
A watermark could give people another clue about where text came from, especially when AI-written material is passed off as human work. Extending it to more models widens the reach of that idea, although it does not make every Claude answer easy to identify by sight. The method described here is subtle by design, not a label sitting politely at the bottom of the page.
Anthropic says the pattern will not affect accuracy or readability. That is the company’s description of the system, not evidence in this report that every output can be reliably identified or that the watermark cannot be removed or imitated.
Our read
Expanding coverage is a practical step towards making AI-generated text easier to trace. But a watermark is only useful if it survives ordinary editing and can be checked reliably. Readers should treat it as a promising provenance signal, not a little digital certificate proving who wrote every sentence.
What to watch
- Whether Anthropic expands the watermark to more Claude versions.
- How the company explains the method and its limits.
- Whether independent testing finds the watermark reliable after text is edited or copied.
Discussion spark: Should AI companies be expected to watermark every model’s output, even if the marks can be removed or cannot prove who wrote a text?
Sources and evidence
- Source update (1 November 2026, 16:59 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.