Claude now watermarks its API text output — no opt-out

Mythos 5, Opus 4.7 and Sonnet 4.6 now embed an invisible mark in every generation. Non-default sampling can drop it, and the detector stays gated.

Nowline NOV 6 3:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Three models, marked at the source

    Text from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 now carries Anthropic's watermark at the model level — present in the Claude API, Claude, Claude Code, Cowork, and Tag no matter the surface. It adds no tokens or hidden characters and no data about you, there's no opt-out, and you change nothing in your requests.

  • The sampling gotcha that silently strips it

    On Sonnet 4.6, setting temperature, top_p, or top_k to a non-default value might not carry the watermark. Anthropic's guidance: if you rely on the mark, omit all three — including any a framework or client library quietly sets for you.

  • AWS, Google Cloud, and Foundry are in scope too

    Supported models served through AWS, Google Cloud, and Microsoft Foundry will also carry the mark, with the Amazon Bedrock rollout finishing by November 10. Routing Claude through a cloud provider won't keep your output unmarked.

  • You can't check the mark yourself yet

    Watermark detection is in private preview, limited to regulators, law enforcement, media, fact-checkers, researchers, educators and EU civil-society groups — plus enterprises verifying their own Claude output. Everyone else applies by form, and Anthropic says access widens over time.

  • Why now, and what survives

    The rollout delivers on Anthropic's signing of the EU AI Act's Article 50(2) transparency code: EU models launched on or after Aug 2, 2026 are marked at launch, with older ones backfilled. The mark rides through copy-paste and light edits, but heavy paraphrase, translation, or very short text can lose it — and a missing mark never proves output wasn't Claude's.