Anthropic Now Watermarks All Claude Output: Text Watermarks and C2PA for Files

Last updated: October 31, 2026 (re-verified against Anthropic's help center: detection is now in private preview, and older models are being added)

Anthropic now marks the text and files its newer Claude models generate, and is extending that to older models. The company updated its support documentation on August 11 to confirm it is rolling out machine-readable watermarking under its commitments to the EU AI Act's Code of Practice on Transparency of AI-Generated Content, and the change applies everywhere Claude is offered, including the API, Claude Code, Cowork, and Tag.

The headline number: all Claude models launched on or after August 2, 2026 support marking at launch, and Anthropic says it is adding watermarks to models released before that date under the law's transition period. As of October 31, Anthropic's help center table shows text watermarks on Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5, and no text watermark yet on Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5 or Haiku 4.5.

Two marking mechanisms, applied at the model level#

Anthropic uses two complementary techniques:

• Embedded watermarks in text. When a supported model generates text, it weaves an imperceptible watermark into the text itself. Anthropic states the mark does not change the meaning, quality, or readability of the response, and because the watermark is part of the text it travels with copy and paste and may persist through some editing. Crucially, the mark is applied at the model level, so it is present no matter which product or surface generated the text.
• Signed provenance metadata on files. For supported file types such as .svg, .png, and .jpg, Claude attaches signed provenance metadata following the C2PA open standard, which lets a verifier detect whether a file was processed by Claude and whether it has been tampered with.

The coverage is broad. Marking applies to output from supported models across the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and it also applies when supported models are accessed through AWS, Google Cloud, or Microsoft Foundry, though Anthropic notes signed provenance metadata may not be supported on every partner platform. The marks ship worldwide, not just in the EU.

Detection is in private preview, with no public text checker#

The missing half of the loop has partly shipped. Anthropic says watermark detection "is currently in private preview, available to eligible organizations as required under EU law," such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus enterprises that must verify watermarking for their own compliance. It plans to expand access over time, and anyone else can register interest through its access request form. There is no public checker for text. For files there is one: the free Claude Content Checker reads the C2PA Content Credential.

Older models are next. Third-party reporting based on an email Anthropic sent Claude Platform customers says four older models (Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6) start carrying the text watermark on November 3, with Amazon Bedrock a few days later (CellCog). We could not find that date on an Anthropic page we could open, so treat it as reported rather than confirmed and check your own Platform email. Anthropic's explainer says only that it is working to add watermarking to older models "over the coming months."

The support page is explicit about what marks are not. A detected mark does not confirm full provenance: Claude is often used to proofread, translate, summarize, or convert files, so output can carry a mark even when the underlying ideas came from elsewhere. And the absence of a mark proves nothing, because the text may have been heavily edited, paraphrased, translated, or mixed with other writing, may be too short for a reliable signal, or may have had its metadata stripped by format conversion, re-saving, or screenshots.

What this means for developers#

This is the most concrete, dated, default-on text watermarking commitment any large model provider has put on the table, and it lands directly in the developer workflow, not just in a consumer chat app.

Code is covered, with limits. Claude Code output is covered by the model-level mark, but Anthropic's explainer says watermarking needs word choices that are equally good, so exact code tokens are generally not marked and comments, commit messages and plan text carry more of it. Even with detection access widened, a "no AI-generated code" policy would rest on a weak signal for code itself. The flip side is the false-positive problem: the support page itself concedes marks are a signal, not conclusive proof, so a detection-based policy will inevitably mislabel human code that merely passed through a Claude edit or review. Teams will need to decide whether that trade is acceptable before wiring detection into CI.

API consumers inherit the mark. The watermark is embedded in raw text responses, and Anthropic claims it survives copy and paste and some editing. Post-processing chains that translate, reformat, or summarize responses can still strip it, per the limitations section, so the watermark's durability in real pipelines is the open question.

EU-facing products built on Claude have their own obligations. Anthropic explicitly tells builders: "you should independently assess what Article 50 requires of your products and services." The company says it will publish technical guidance on its marking and detection approach, so a developer shipping a Claude-backed product in the EU still has their own assessment to do.

C2PA is the piece to integrate early. Signed provenance metadata on generated images and SVG files is standard-compliant and verifiable, so any pipeline that stores or distributes Claude-generated assets can validate signatures now with the free Claude Content Checker, without waiting for wider text-detection access.

Our take#

The model-level, default-on design is the right architecture, and it is more honest than the file-level metadata approach alone, which dies on the first screenshot. The two open questions are durability and detection. Anthropic says light editing probably leaves the mark and a full rewrite removes it, but there is no independent measurement yet, and detection is limited to approved organizations. Treat the announcement as a compliance commitment with a narrow verification story, and treat any tool that claims to detect Claude text without Anthropic's key as unverified.

The move also widens the gap between providers. Google and OpenAI have published provenance commitments of their own, but Anthropic is the one that has now named a concrete, dated mechanism for text. That is the pattern to watch: watermarking stops being a research demo and becomes a default property of an API contract, which changes how attribution, review tooling, and compliance checkboxes work for everyone building on top.

FAQ#

Which Claude models watermark their output?#

All Claude models launched on or after August 2, 2026 support marking at launch. Anthropic's help center lists text watermarks on Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5, and says it is adding them to earlier models. Reports say Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 start on November 3. Marking applies across the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry for supported models.

Is code generated by Claude Code watermarked?#

Yes. The text watermark is applied at the model level, so it is present no matter which product generated the text, and Claude Code output is covered. That includes generated code, commit messages, and plan text.

Can I detect a Claude watermark today?#

Only if you are an approved organization. Text watermark detection is in private preview for regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil society groups and enterprises with matching compliance duties. Anyone can check files for a C2PA Content Credential with the free Claude Content Checker. Treat any other tool that claims to detect Claude text as unverified.

Does a missing watermark prove text was not written by Claude?#

No. Per Anthropic's support page, the text may have been heavily edited, paraphrased, translated, or mixed with other writing, may be too short for a reliable signal, or may have had its metadata stripped by format conversion, re-saving, or screenshots. A detected mark is also only a signal that content may have been processed by Claude, not proof of full provenance.

Continue Reading#

• VS Code Copilot Co-Author Attribution - why attribution of AI-generated code is really a consent and audit problem
• Ghost Font: Text That Humans Can Read But AI Cannot - the other side of the coin, hiding text from models instead of marking it
• Apertus: Europe's Answer to AI Sovereignty - how EU compliance shapes open model development
• Agent Identity as a Security Layer - provenance and identity for agent output in production workflows
• Claude Code Permissions Settings Guide - how Claude Code handles boundaries before it touches your repo
• The Exponential and the Working Developer: Sitting With Amodei's Hardest Questions

Sources#

• Anthropic explainer, "How Claude's text watermarking works": anthropic.com/news/claude-text-watermark
• Anthropic support article, "How Claude marks AI-generated content" (re-read October 31, 2026): support.claude.com/en/articles/16266773
• European Commission, "Drawing-up a General-Purpose AI Code of Practice": digital-strategy.ec.europa.eu/en/policies/ai-code-practice
• C2PA specification: c2pa.org