Which Claude Models Have Watermarks? A Current Coverage Guide

Short answer: Do not determine Claude watermark coverage from a plan name such as Free, Pro, Max, Team, or API. Anthropic describes marking as a model-level capability. Its current guidance says supported models launched on or after August 2, 2026 include machine-readable marking from launch, while work on earlier models is ongoing. The model and output type therefore matter more than whether the user paid for the account.

The word supported is important. Anthropic has not published a permanent model-by-model verification table or an open detector. Coverage can change as models are introduced, retired, or updated, so a current model ID and the latest provider documentation are better evidence than a screenshot of a subscription page.

Coverage at a glance

Situation	Best current interpretation
Supported Claude model launched on or after August 2, 2026	Marking is described as available from launch.
Claude model released before that date	Anthropic says earlier models are being updated; do not assume uniform coverage.
Free account versus paid account	No public evidence shows that payment tier is the controlling factor; marking is described at model level.
Claude through an API or another Claude surface	If the same supported model generates the output, the product surface alone should not be treated as an exemption.
Standard editing that does not substantially change input	EU guidance describes an exception, but the exact product behavior still requires provider confirmation.
Source code	European Commission guidance lists source code outside the Article 50 marking obligation; that legal scope is not a model-by-model product guarantee.
Supported generated file	It may carry signed provenance metadata rather than relying only on the text signal.

Why August 2, 2026 matters

Article 50(2) of the EU AI Act requires providers of systems that generate synthetic text, image, audio, and video content to mark output in a machine-readable format and make it detectable as artificially generated or manipulated, as far as technically feasible.

The European Commission's Article 50 FAQ explains several scope limits. Short sequences, source code, machine-to-machine output, and some closed-loop development uses can fall outside the marking obligation. Standard editing assistance that does not substantially alter the input or its meaning is also treated differently.

These legal categories help explain a rollout, but they should not be used to invent product behavior. Anthropic's own documentation remains the primary source for whether a particular Claude model and output are supported.

New models versus earlier models

Anthropic's marking guidance describes a forward-looking rule: supported models launched on or after the legal deadline include marking at launch. Earlier models are not automatically equivalent. Anthropic says it is working to extend support, which means coverage during the transition can be uneven.

For a user, this produces a simple verification sequence:

1. Record the exact model name or API model ID.
2. Record when the output was generated.
3. Check whether the model launched before or after August 2, 2026.
4. Review Anthropic's latest support page for the word supported and any exclusions.
5. Record whether the output was plain text, source code, an edited input, or a generated file.

Without those details, a statement such as “all Claude text is watermarked” is too broad. The opposite statement—“my old Claude plan has no watermark”—is equally unsafe.

Does the free or paid plan change watermarking?

Searches frequently ask whether Claude Free and Claude Pro use different marks. Anthropic describes the mark as operating at model level and across the places where a supported model is offered. The public guidance does not identify subscription price as the deciding factor.

Plans can still affect which models and features are available. If a Free user and an API developer receive output from different models, their practical coverage may differ because of the model, not because one account is free and the other is paid.

The best answer is therefore conditional: identify the actual model. Do not infer watermark status from the billing label alone.

What about Claude, Claude Code, and the API?

A model-level system is intended to avoid a gap in which the same model marks output in one interface but not another. Anthropic's public explanation says marking follows supported model output across Claude surfaces. Nevertheless, output context matters.

• Ordinary natural-language text is the clearest case described by the text-marking guidance.
• Source code is listed by the European Commission as outside the Article 50 marking obligation. This does not prove that no Claude product ever attaches other provenance or metadata to code-related files.
• Files created by Claude can use signed provenance metadata when the format and workflow are supported.
• Editing assistance may fall within a legal exception when it is standard editing and does not substantially change the input. The boundary depends on what the system actually did.
• Third-party apps using Claude should identify their model and output pipeline. A product name alone may conceal which provider, model version, or post-processing steps produced the final text.

Can you test whether a specific model marked a passage?

Not reliably with an ordinary AI writing detector. A style classifier estimates resemblance. A Unicode scanner reports particular characters. Neither becomes an official watermark verifier merely because its result mentions Claude.

Anthropic says third-party detection support and more detailed technical guidance are being developed. Until a compatible verification mechanism is available, the responsible description is limited: a tool may identify style overlap or content-hygiene findings, but it should not certify a provider-issued mark.

The FaddyAI Claude Watermark Detector follows that distinction. It highlights the exact style and Unicode findings used in its result and states that it is not an official Anthropic detector.

How to maintain an internal coverage record

Organizations using Claude in publishing, education, software development, or regulated work should maintain a small evidence table rather than relying on memory. Record:

• model name and version;
• access surface, such as Claude chat or API;
• generation date;
• output type;
• degree of human editing;
• applicable disclosure policy;
• link and review date for Anthropic's current guidance.

Review the table when Anthropic launches a model or publishes technical verification details. This record is more defensible than assuming every document made with “Claude” has the same provenance behavior.

The practical conclusion

Current Claude watermark coverage is model-dependent and still evolving. Supported models launched on or after August 2, 2026 are the clearest covered group in Anthropic's guidance. Earlier models require current confirmation. Free versus paid access is not the published dividing line, and special cases such as editing assistance, source code, and generated files need separate treatment.

When the exact model is unknown, say that coverage cannot be determined from the plan or interface alone.

Primary sources reviewed

• Anthropic: How Claude marks AI-generated content
• Anthropic Transparency Hub
• European Commission: Article 50 transparency FAQ
• EU AI Act, Article 50

Related FaddyAI resources

• FaddyAI Claude Watermark Detector