Guides
Which Claude Models Have Watermarks?
Which Claude models have watermarks comes down to a launch date rather than a product list.
Anthropic's rule is that models launched on or after 2 August 2026 support machine-readable marking at launch, with earlier models described as in progress.
Launched on or after 2 August 2026: marked. Earlier models: being brought into scope.
Runs in your browser. No signup, no upload, nothing stored.
The rule as stated
Marking support ships with models launched from 2 August 2026 onward. Reporting has named Claude Sonnet 4.6 and Claude Haiku 4.5 among those covered at launch.
Models predating the rollout are described as in progress, which means the boundary moves over time rather than staying fixed. A model unmarked when you checked may not stay that way.
Every surface is included
Where you access the model does not change the answer at all:
- Claude Platform (API)
- Claude
- Claude Code
- Claude Cowork
- Claude Tag
- AWS, Google Cloud and Microsoft Foundry deployments
Why 'in progress' matters for archives
Text generated before your model was brought into scope was not marked at generation time, and cannot be marked retroactively.
The mark is applied during generation, as the tokens are chosen. Text already written is fixed — nothing Anthropic ships later reaches back into a document sitting on your disk.
Pinning an old model is a shrinking strategy
It works, narrowly, while an unmarked model remains available to you. It also means giving up every capability improvement in exchange.
Coverage is expanding by design and there is no opt-out. Anyone building a workflow on the assumption that an unmarked model stays available is building on a deprecation schedule.
How to check which model you used
In the Claude interface the model is shown in the conversation; through the API it is whatever string you sent in the request.
If you are working from a saved document with no record of either, the file's creation date is usually enough on its own. Anything written before 2 August 2026 predates the rollout entirely.
The practical default
Assume anything you generate from now on is marked, and plan for that rather than for the exception.
It is the assumption that stays correct as the boundary moves, and it costs nothing to hold. Every alternative requires tracking a moving target for a benefit nobody can currently verify anyway.
The one case where the distinction genuinely matters is an archive: work generated well before August 2026 was never marked, and nothing shipped since reaches back into it.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.