Claude's invisible watermark now reaches API and Claude Code output

The SynthID-style mark is on by default for Opus 4.8, 4.7 and Sonnet 4.6 with no opt-out; code is largely spared and detection stays locked to regulators.

Nowline NOV 4 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your Claude output now ships with an invisible mark

    Anthropic has extended its SynthID-style text watermark to its current production models — Opus 4.8, Opus 4.7 and Sonnet 4.6 — so output served through the Claude API, Claude Code, Cowork and the apps now carries it worldwide, with no opt-out on any tier, free or enterprise.

  • Pin an older model if you need clean text

    The mark is live on the whole 5.x line plus Opus 4.8/4.7 and Sonnet 4.6, but Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carry no text watermark — pinning one of those is the only way to get unmarked prose today. The Bedrock rollout for the newer models completes Nov 10.

  • Your code comes out almost clean

    Where an exact output is required, Anthropic skips the watermark, so generated code is essentially unmarked — only free choices like comments or variable names can carry a faint signal, with negligible effect on the code. Prose, docs and commit messages are fully marked.

  • No token tax, and it survives copy-paste

    Anthropic says the mark adds no tokens, no measurable latency and no price change, and doesn't touch quality. It rides through copy-paste and light edits; a full rewrite, paraphrase, translation or very short snippet drops it. Sonnet 4.6 may miss it at non-default temperature, top_p or top_k.

  • You can't run the detector on your own text

    Detection is a private-preview API gated to EU-obligated groups — regulators, law enforcement, media, researchers, educational and civil-society orgs — plus compliance-bound enterprises; ordinary API customers get nothing. And a hit only says Claude was 'likely involved,' never that it authored the text.