Which Claude models are watermarked? (live list)
By Claude Watermark Research · Updated
As of 2 September 2026, Claude Fable 5.1 and Claude Mythos 5.1 carry Anthropic's statistical text watermark. Both launched on 1 September 2026, after the 2 August 2026 marking cutoff. Opus 5, Sonnet 5, Fable 5, Mythos 5, Opus 4.8 and Haiku 4.5 launched before that date and remain unmarked while Anthropic works through its transition period. This page is dated because Anthropic can add marking support to older models.
Check and remove AI traces from my text. Checking runs in your browser, no account, nothing uploaded.
The rule is a launch date, not a model family
Anthropic's condition is about when a model launched, not what it is called or how capable it is. Any model that launches on or after 2 August 2026 is in scope. Anything that shipped before is not, unless Anthropic retrofits it, which it has said it is working towards but has not done.
This is why the question “is Claude watermarked?” has no single answer. Fable 5.1 is marked, while Fable 5 is not. The exact model and launch date decide the status.
Why the model name matters
Fable 5 and Fable 5.1 sound like minor variants, but they sit on opposite sides of Anthropic's cutoff. Fable 5 launched before 2 August and is not marked; Fable 5.1 launched on 1 September and is marked from day one.
The confusion is compounded by a real artifact that does exist today: copying from the Claude web interface carries HTML class names such as `font-claude-response-body` into whatever you paste into. That is genuinely detectable, it is what the AO3 skin catches, and it has nothing to do with the statistical watermark.
What a marked model changes
The mark is distributional: a secret key biases the choice among near-equivalent next tokens, leaving a statistical signature spread across the passage. It survives copy-paste because the signal is the word choice itself.
Detection requires Anthropic's key. Its detector is in private preview for eligible organisations, but no public checker can confirm a passage is marked or certify a removal. Any public tool claiming verified detection should disclose the access it relies on.
Questions
- Is Claude Fable 5.1 watermarked?
- Yes. Claude Fable 5.1 launched on 1 September 2026 and Anthropic explicitly lists it as marked at launch.
- Is Claude Opus 5 watermarked?
- No. Opus 5, Sonnet 5, Fable 5, Mythos 5, Opus 4.8 and Haiku 4.5 all launched before 2 August 2026, so they fall outside the policy.
- Will older models be watermarked retroactively?
- Anthropic has said it is working to extend marking to earlier models. Until that ships, output from those models carries no statistical mark.
- Does the API differ from claude.ai?
- No. The policy is model-based and applies across claude.ai, the API, Claude Code and the other surfaces. There is no paid tier that disables it.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Check and remove AI traces
