News

Claude's watermark covers models launched from 2 August — Anthropic has not launched one since

On 11 August Anthropic published how watermarking works in Claude: the model embeds an imperceptible pattern in generated text by biasing what the company calls "low-stakes choices" between equally good words. The mark survives copying and pasting, is invisible to a reader and is meant to satisfy Article 50 of the EU AI Act, which became enforceable for newly launched systems on 2 August 2026. Anthropic applies it worldwide, saying it has no durable way to scope the behaviour by region. The scope is narrower than the coverage suggests. Anthropic's own wording is that "Claude models launched on or after August 2, 2026 will support machine-readable marking at launch", and that the EU law grants a transition period for models launched before that date, to which the company is "working to add" watermarking over the coming months. No completion date is given. Against our catalogue that rule currently applies to nothing. We hold twenty Claude profiles, twelve of them still in service. The most recent release is Claude Opus 5 on 24 July 2026 — nine days before the cutoff. Claude Sonnet 5 arrived on 30 June, Fable 5 and Mythos 5 on 9 June, Opus 4.8 on 28 May. Every model a customer can call today therefore falls on the transition side of the line, including the ones running on AWS, Google Cloud and Microsoft Foundry, where Anthropic notes coverage may be further limited. Anthropic is unusually plain about what the mark does not do. Detection "doesn't work well on small samples", the signal is sparser in factual passages where fewer word choices are available without hurting accuracy, and code carries less watermarking than prose. If Claude only edits or proofreads a human text, the mark may be undetectable; a complete rewrite of every word removes it. The company also stresses that finding a watermark proves the text passed through Claude, not that Claude wrote it. Images and files are handled separately, through C2PA content credentials in file metadata, which a screenshot or a format conversion strips. The detection tool that would let anyone check a passage is not out yet either — Anthropic says an API is coming, with technical documentation to follow. For readers the practical reading is this: from today's models, an unmarked Claude answer is the normal case, not evidence of tampering. That will only change when the first post-cutoff model ships, or when the retrofit reaches the current family.