Most drafts on this blog start in Claude, so what I publish carries a statistical fingerprint I can’t switch off. I’d rather disclose model help plainly than let a watermark do it for me, especially when nobody outside a private preview can run the detector.
For builders, the limits matter more. A hit means Claude was probably involved somewhere, including translating or editing your own words. Paraphrase, mix or shorten the text and the signal fades. So no gate in my pipelines will read ‘watermark’ as guilt or ‘no watermark’ as human. It’s a hint, not a verdict.
The story — Anthropic is extending its invisible text watermark to Claude Fable 5, Sonnet 5 and Opus 4.8 from 30 September, per a customer email reported by The Decoder. Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5 already carry it, and older models follow by 2 December. Users can’t disable it. The EU AI Act prompted it, but it applies worldwide. It’s a SynthID-Text variant that nudges token choice into a pattern detectable with a key. The detector API remains in private preview. (Source)