Remove Claude Opus 5 watermark

Opus 5 shipped after Anthropic's 2 August 2026 cutover, so everything it writes carries the embedded mark. That covers the Claude Platform API, claude.ai, Claude Code, and Claude models served on AWS, Google Cloud and Microsoft Foundry. Anthropic's own framing is that the mark proves the text passed through Claude, not that a human did or did not write it.

Zero-width characters

Invisible code points (U+200B–U+200D, U+2060, U+FEFF) interleaved between words. They survive copy-paste and carry a machine-readable payload while rendering as nothing.

Variation selectors

U+FE00–U+FE0F and the U+E0100–U+E01EF supplement attach to a preceding character without changing how it looks, giving a high-bandwidth channel to hide bits in plain text.

Non-breaking and exotic spaces

U+00A0, U+2009 thin space, U+202F narrow no-break space and friends used in place of a plain space to encode a signal in the gaps.

Statistical (token-level) watermarks

A cryptographic key biases the model's token sampling toward a pseudo-random 'green list'. Nothing is added to the text. The distribution itself is the mark, so only rewriting enough of the wording removes it.

File and document metadata

Provenance fields written into DOCX/PDF/image containers (C2PA manifests, XMP, EXIF) alongside the visible content.

Paste it above. If Claude Opus 5 text is in there, the report names every carrier it found with its code point, and the cleaned copy appears next to it.

For this model the findings are usually zero-width characters, variation selectors, non-breaking and exotic spaces.

Anthropic announced on 11 August 2026 that every model released after 2 August automatically embeds an imperceptible watermark in generated text and files, to satisfy Article 50 of the EU AI Act, and applied it globally rather than only in the EU. The mark travels with a copy-paste and can survive light editing.

There are two different things people mean by an AI watermark, and only one of them is something a cleaner can touch.

The first lives in the characters: invisible code points, look-alike letters, unusual spaces. Those are data added to the string, and removing them is exact and verifiable. The report tells you precisely what came out.

The second is statistical. A key biases which words the model picks, so the mark is spread across the word choices themselves and nothing was inserted at all. No character-level tool can remove that, including this one. Only rewriting enough of the wording does, and any site claiming otherwise is selling you something. We would rather say so than let you publish on a false assumption.

1. Paste the textPaste the Claude Opus 5 output into the box at the top of this page. It stays in your browser tab; the scan runs against the characters, not against any account you have.
2. Read the reportEvery flagged character is listed with its Unicode name and code point, so you can verify the finding in any hex editor rather than taking our word for it.
3. Copy the clean textThe cleaned copy is byte-identical to what you pasted apart from the characters listed in the report. Nothing is rephrased, and nothing is sent anywhere.

Does Claude Opus 5 watermark its text?

Confirmed watermarking. Opus 5 shipped after Anthropic's 2 August 2026 cutover, so everything it writes carries the embedded mark. That covers the Claude Platform API, claude.ai, Claude Code, and Claude models served on AWS, Google Cloud and Microsoft Foundry. Anthropic's own framing is that the mark proves the text passed through Claude, not that a human did or did not write it.

Is removing a Claude Opus 5 watermark legal?

Stripping invisible characters from text you own is ordinary text processing, the same thing your word processor does when it normalises a paste. What matters is what you do next: passing AI-written work off as your own where honesty is required (academic submissions, disclosure-bound contexts, contracts that say otherwise) is a problem regardless of what any tool did to the characters.

Will this get Claude Opus 5 text past an AI detector?

Probably not, and we are not going to pretend otherwise. Detectors for this family of models score word choice and sentence structure, not characters. Cleaning changes the bytes; it does not change the writing.

Do you store what I paste?

No. The scan records counts and code-point classes for our own usage numbers and nothing else. Not the text, not a fragment of it, not a hash of it.

Does the cleaned text still mean the same thing?

Yes. Deletions are limited to characters that render as nothing, and substitutions map to the ASCII character that looks identical. Joiners that carry meaning, inside emoji sequences or in scripts where a joiner is part of the spelling, are reported and deliberately left in place.