Anthropic watermark detection API status

What this means

Anthropic has announced that it plans to provide a watermark detection API, but its public documentation, pricing, access rules and response format have not yet been released. No independent website can currently verify Claude's private watermark using Anthropic's detector.

The primary source is Anthropic's own announcement: Anthropic — Watermarking Claude's text output. Anything on this page that is not in that announcement is our own integration status, and is labelled as such.

Published by Anthropic

• Claude's text output is watermarked at the model level.
• The watermark is statistical, based on word and token selection driven by a keyed source of randomness.
• Nothing is added to the text; there are no hidden characters.
• A watermarked response is indistinguishable from an unwatermarked one to a reader.
• A watermark detection API is planned, with implementation details still being worked out.

Not yet published

• Endpoint, request shape and response schema.
• Access rules, pricing and rate limits.
• Minimum reliable sample size and how confidence is reported.
• Whether the terms permit iterative use — re-checking edited text — which is what an API-guided editing workflow would require.
• Data handling, retention and logging for submitted text.

FreeAIText's integration state

Built and waiting, not connected:

• A backend-only provider interface with a normalisation layer, so the interface does not depend on Anthropic's future response schema.
• A local estimator provider — what runs today, labelled a research estimate wherever it appears.
• An Anthropic provider that is deliberately inert, with no endpoint. Guessing one before Anthropic publishes it would be inventing a fact.
• An explicit feature flag. Setting environment variables is not enough on its own; the flag must be switched on, and any failure falls back to the clearly labelled local estimator rather than presenting a local score as an Anthropic result.
• Distinct badges for local estimates and provider results, so the two cannot be confused.
• An API-guided re-check workflow that stays disabled unless Anthropic's terms explicitly permit iterative verification.

Longer write-up: what is known about Anthropic's watermark detection API so far.