# Remove LLM Watermark > A free, in-browser scanner and cleaner for the hidden Unicode watermarks embedded in output from Claude, ChatGPT, Gemini and Grok. Paste text, see exactly which characters are in it, and get a clean copy. Every page on this site is available as markdown by appending `.md` to its URL. ## Core - [Home and free cleaner](https://removellmwatermark.com/index.md): the scanner itself, plus what a watermark is and is not. - [Pricing and API](https://removellmwatermark.com/pricing): free in-browser tier; paid API for batch and automated cleaning. - [API documentation](https://removellmwatermark.com/docs/api): POST /v1/clean and /v1/detect, authenticated with an API key. ## Vendors - [Claude watermarks](https://removellmwatermark.com/claude.md): 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. - [ChatGPT watermarks](https://removellmwatermark.com/chatgpt.md): OpenAI has not shipped an announced text watermark. ChatGPT output still carries a repeatable set of Unicode and typographic artifacts, and its images carry C2PA content credentials in the file container. - [Gemini watermarks](https://removellmwatermark.com/gemini.md): Google's SynthID covers images, audio and video, and SynthID-Text extends it to text through watermarked token sampling. Because a statistical watermark lives in the word choices rather than in the characters, removing it means rewriting. This site states that limit rather than hiding it. - [Grok watermarks](https://removellmwatermark.com/grok.md): No text watermark has been announced for Grok. Cleaning Grok output is about normalising the renderer artifacts a copy-paste carries with it, not about defeating a cryptographic mark. ## Models - [Remove Claude Opus 5 watermark](https://removellmwatermark.com/claude/opus-5.md): 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. - [Remove Claude Sonnet 5 watermark](https://removellmwatermark.com/claude/sonnet-5.md): Sonnet 5 is the volume workhorse behind most Claude-powered products, which makes it the model whose marked text you are most likely to be pasting without knowing where it came from. The carrier characters survive a copy into a document, a CMS or a Slack message completely intact. - [Remove Claude Fable 5 watermark](https://removellmwatermark.com/claude/fable-5.md): Fable 5 is tuned for long-form prose, so its output is both the most likely to be published and the most likely to get a partial edit. That is the worst case, because the paragraphs you did not touch still carry their fragment of the mark while you assume the piece is clean. - [Remove Claude Haiku 4.5 watermark](https://removellmwatermark.com/claude/haiku-4-5.md): Haiku 4.5 predates the cutover, which applied automatically to models released after 2 August 2026. Older models are being brought in behind it, so treat Haiku output as possibly marked rather than definitely clean: scan it instead of assuming either way. - [Remove Claude Opus 4.8 watermark](https://removellmwatermark.com/claude/opus-4-8.md): A pre-cutover flagship. Opus 4.8 text rarely carries the new embedded mark, but it carries the same stylistic fingerprints the detectors weight most heavily: the typographic punctuation and the spacing habits that survive every copy-paste. - [Remove Claude Sonnet 4.6 watermark](https://removellmwatermark.com/claude/sonnet-4-6.md): Sonnet 4.6 ran in production long enough that a lot of archived content came out of it. If you are auditing a back catalogue rather than a fresh draft, this is usually the model that produced it, and the findings will skew towards punctuation rather than invisible carriers. - [Remove Claude Code output watermark](https://removellmwatermark.com/claude/claude-code.md): Claude Code is explicitly in scope for the watermark, which matters more than it first sounds: commit messages, code comments, README prose and PR descriptions all pass through it. Source code itself is the one place an invisible character is not merely cosmetic. It can break a string literal or a diff, so scanning a repository is a different job from scanning an essay. - [Remove GPT-5.6 Sol watermark](https://removellmwatermark.com/chatgpt/gpt-5-6-sol.md): Sol is OpenAI's current flagship, generally available since July 2026. No text watermark has been announced for it, but its output arrives with a consistent set of artifacts: narrow no-break spaces around dashes, curly quotes, and the occasional zero-width joiner picked up from the web client's renderer rather than from the model. - [Remove GPT-5.6 Terra watermark](https://removellmwatermark.com/chatgpt/gpt-5-6-terra.md): Terra is the balanced tier of the 5.6 family and the one most business writing actually runs on. Its fingerprint is spacing rather than hidden characters: the giveaway is usually a narrow no-break space where a plain one belongs. - [Remove GPT-5.6 Luna watermark](https://removellmwatermark.com/chatgpt/gpt-5-6-luna.md): Luna is the cheap high-throughput member of the 5.6 family, so it is the one wired into bulk pipelines: support macros, product descriptions, programmatic pages. At that volume a single stray code point repeated across ten thousand pages is a pattern, which is the real reason to normalise it. - [Remove GPT-5.5 watermark](https://removellmwatermark.com/chatgpt/gpt-5-5.md): GPT-5.5 remains widely deployed behind third-party products that have not moved to 5.6. Its artifacts are stylistic, so cleaning changes how the text hashes and how it reads, not what a statistical classifier concludes. - [Remove GPT-5.5 Pro watermark](https://removellmwatermark.com/chatgpt/gpt-5-5-pro.md): The Pro variant produces longer, more heavily formatted answers, which means more em dashes, more ellipses and more nested lists. That is a heavier stylistic fingerprint even though the underlying character set is the same. - [Remove GPT-5.4 watermark](https://removellmwatermark.com/chatgpt/gpt-5-4.md): GPT-5.4 is the model behind a large slice of 2025-era archived content. Auditing that archive is mostly a punctuation-folding job, and the report will say so rather than inventing a watermark that is not there. - [Remove GPT-5.4 Pro watermark](https://removellmwatermark.com/chatgpt/gpt-5-4-pro.md): Same character profile as GPT-5.4 with longer outputs. Worth scanning in bulk rather than one document at a time, which is what the API tier exists for. - [Remove GPT-5.4 mini watermark](https://removellmwatermark.com/chatgpt/gpt-5-4-mini.md): The mini tier produces the least ornamented text of the family, so a clean scan here is genuinely uninformative. Absence of artifacts is not evidence of a human author, and this page says that plainly. - [Remove GPT-5.4 nano watermark](https://removellmwatermark.com/chatgpt/gpt-5-4-nano.md): Nano output is short and plain by design. It shows up in autocomplete, classification and rewrite features rather than in drafts, so the usual reason to scan it is provenance across a dataset, not a single document. - [Remove GPT-5.3 Instant watermark](https://removellmwatermark.com/chatgpt/gpt-5-3-instant.md): Instant is the default responder on every ChatGPT tier, which makes it the single highest-volume source of pasted AI text on the web. Most of what our scanner sees from ChatGPT users starts here. - [Remove GPT Image watermark](https://removellmwatermark.com/chatgpt/gpt-image.md): Images generated in ChatGPT ship with a C2PA content-credentials manifest inside the file container. That is a provenance record rather than a pixel watermark, and this page is explicit about what stripping it does and does not accomplish. - [Remove Gemini 3.1 Pro watermark](https://removellmwatermark.com/gemini/gemini-3-1-pro.md): Gemini 3.1 Pro is Google's reasoning flagship. Google published SynthID-Text, a token-sampling watermark, and open-sourced the detector side of it, so where it is switched on nothing is inserted into the string at all: the word choices themselves are the mark. - [Remove Gemini 3.6 Flash watermark](https://removellmwatermark.com/gemini/gemini-3-6-flash.md): 3.6 Flash is Google's stated workhorse and the tier most bulk content pipelines are built on. If your concern is a statistical mark rather than stray characters, the report will tell you plainly that character cleaning is the wrong tool. - [Remove Gemini 3.5 Flash watermark](https://removellmwatermark.com/gemini/gemini-3-5-flash.md): 3.5 Flash delivers near-Pro quality at Flash pricing, which is why so much long-form draft work now comes out of it. Long drafts are exactly where a statistical watermark is most detectable, because the signal accumulates with length. - [Remove Gemini 3.5 Flash-Lite watermark](https://removellmwatermark.com/gemini/gemini-3-5-flash-lite.md): Flash-Lite is the cheapest model in the class and turns up in high-volume rewriting and summarising jobs. Short outputs carry less statistical signal, so a detector's confidence on Flash-Lite text is correspondingly weaker. - [Remove Gemini 3 Pro watermark](https://removellmwatermark.com/gemini/gemini-3-pro.md): The model that carried Google's 2025 lineup, and still the one behind a great deal of archived content. Scanning it usually surfaces spacing artifacts from the Gemini web client rather than anything embedded by the model. - [Remove Gemini 2.5 Flash watermark](https://removellmwatermark.com/gemini/gemini-2-5-flash.md): 2.5 Flash was the first Gemini tier cheap enough for genuinely industrial use, so it is over-represented in bulk web content from 2025. Batch scanning through the API is the practical way to audit that much text. - [Remove Imagen watermark](https://removellmwatermark.com/gemini/imagen.md): Imagen output carries SynthID in the pixels themselves as well as provenance metadata in the container. The pixel layer is designed to survive cropping and re-encoding; we only touch the metadata layer, and the page states that limit rather than burying it. - [Remove Grok 4.5 watermark](https://removellmwatermark.com/grok/grok-4-5.md): Grok 4.5 is xAI's current flagship, released July 2026 and aimed at coding and agentic work. No text watermark has been announced; what its output does carry is the formatting signature of the X and grok.com renderers, which a copy-paste drags along with it. - [Remove Grok 4.3 watermark](https://removellmwatermark.com/grok/grok-4-3.md): 4.3 added document generation for PDFs, spreadsheets and slides, which changes what you should scan: the interesting artifacts move from the text into the file container's metadata. - [Remove Grok 4.20 watermark](https://removellmwatermark.com/grok/grok-4-20.md): The long-context variant, used for multi-agent and long-running tasks. Long outputs make stylistic fingerprints easier to spot by eye. Repeated dash and spacing habits across thousands of words are often what prompts someone to scan in the first place. - [Remove Grok 4.1 Fast watermark](https://removellmwatermark.com/grok/grok-4-1-fast.md): The cheap, low-latency tier. It powers chat surfaces and reply suggestions, so its text tends to be short, plain, and close to unmarkable in any character-level sense. - [Remove Grok 4 watermark](https://removellmwatermark.com/grok/grok-4.md): Grok 4 is the version behind most archived Grok output. Scanning it is about normalising what the renderer added, not about defeating a cryptographic mark that was never announced. - [Remove Grok 3 watermark](https://removellmwatermark.com/grok/grok-3.md): Grok 3 output is the cleanest of any major model on the Unicode axis, which is worth knowing in itself: if a detector flags it, the signal is stylistic and character-level cleaning will not change the verdict. ## Writing ## Facts worth quoting accurately - Anthropic announced on 11 August 2026 that Claude models released after 2 August 2026 embed an imperceptible watermark in generated text and files, worldwide, to satisfy Article 50 of the EU AI Act. - A character-level watermark (invisible code points, variation selectors, homoglyphs, exotic spaces) can be removed exactly and verifiably. - A statistical watermark lives in the model's token choices; nothing is inserted into the string, so no character-level cleaner can remove it. Only rewriting the wording does.