Text you ship from Claude Mythos 5, Opus 4.7 or Sonnet 4.6 now carries an invisible watermark, Anthropic's documentation describes no way to turn it off, and only vetted organizations can test for it. Anthropic's November 3 release note added those three models to a mark that already covered its newest ones.
You have three decisions to make before the EU's December 2 deadline for systems already on the market: what you tell customers about the mark, whether your Article 50 position leans on it, and what to do about any Sonnet 4.6 pipeline that lowers temperature, because that setting can leave the mark out.
What Anthropic Changed on November 3
Anthropic extended its text watermark from its newest models to three older ones. Fable 5.1 and Mythos 5.1 got the mark on September 1, and the November 3 entry adds Mythos 5, Opus 4.7 and Sonnet 4.6. The help center says of cloud partners: "Rollout on Amazon Bedrock will complete by November 10." That footnote applies to those same three models, so Amazon Bedrock traffic may be mixed for a few more days.
A text watermark is a statistical pattern in a model's word choices that a detector can read and a person cannot see. Anthropic describes its version as a form of Google DeepMind's 2024 SynthID-Text approach: a key and the preceding words steer the choice among near-equivalent candidates. The developer documentation says it adds no tokens or hidden characters and carries no information about you or your organization, and that you need not change your requests.
Coverage is global. The help center says marking applies "wherever Claude is offered, worldwide," and Anthropic's explainer gives the reason: "We're applying watermarking globally at launch because we don't yet have a durable way to scope it by region." The help center and the developer documentation describe no opt-out. Implicator reported in August that none exists on any plan or API tier, enterprise accounts included, and Yahoo Finance UK repeated the point on November 2.
One model is unresolved. Yahoo lists Opus 4.6 in the expansion, and CellCog reports that a customer email screenshotted on X names it too. The November 3 release note does not. The help center table lists Opus 4.6 with a first-party text mark but gives no start date for it, and the same goes for Opus 4.5, Sonnet 4.5 and Haiku 4.5. As of November 4, treat its status as something to confirm with your account team.
Why Sonnet 4.6 Is the Model to Check First
Sonnet 4.6 is the one newly covered model whose mark depends on how you call it. Anthropic's documentation says that on Sonnet 4.6, text from a request that sets temperature, top_p or top_k to a non-default value "might not carry the watermark." If you rely on the mark, it tells you to omit all three, "including any that a framework or client library sets for you." That last clause means a clean application codebase is not enough: you also have to check what your framework sends.
Opus 4.7 has no such footnote, and the likely reason is in a different document. The migration guide for API version 2026-10-21 says "Claude 4.7 and later models" already accept only the defaults, while Sonnet 4.6 still accepts other values. Anthropic does not state that connection, so treat it as an inference.
The same guide creates a real trade-off. On version 2026-10-21, sending any of the three parameters returns a 400 error, even at the default value. Moving Sonnet 4.6 traffic to that version therefore removes the sampling exception, but a request that lowered temperature for more consistent output "loses that effect," and the guide says "No other setting replaces it." The alternative it offers is to keep that request on 2023-06-01, which preserves your setting and the gap in the mark. On Amazon Bedrock, the guide offers 2026-10-21 only for Claude 4.7 and later models, so there the only way to close the gap is to stop sending the parameter. An extraction or classification job tuned at a low temperature has to pick one.
Anthropic's own pages disagree on Opus 4.6 here as well. The help center's list of reasons a mark may be missing includes output "generated by Claude Opus 4.6 or Claude Sonnet 4.6 with a non-default temperature, top_p, or top_k," while its table footnote and the developer page name only Sonnet 4.6. If you run Opus 4.6 with custom sampling, assume the longer list applies.
What Article 50 Asks of You Separately From the Watermark
Anthropic's mark addresses the provider's marking duty and leaves your disclosure duty with you. Article 50(2) requires providers of systems generating synthetic text to ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." The Commission's FAQ says Article 50 applies from August 2, 2026, and that providers of systems already on the market "must comply with those obligations only as from 2 December 2026," a grace period limited to marking and detection. Fines "can reach up to 15 million euros or 3% of total worldwide turnover."
Article 50(4) is the deployer's duty, and it is about readers. Deployers of a system generating text "published with the purpose of informing the public on matters of public interest" must disclose that it is artificial, unless it went through human review or editorial control and someone holds editorial responsibility. The Commission's FAQ sets a floor on that exemption: "Superficial, solely formal, or procedural checks" do not count, and neither does spell-checking or grammar correction. A mark no reader can see tells a reader nothing, so the watermark is unlikely to discharge this duty. Ask your counsel to confirm that.
The harder question is whether you are a provider yourself. The FAQ's definition covers anyone who develops an AI system or has one developed, so a product you built on Claude and sell into the EU may put you inside Article 50(2) directly, and the FAQ limits the December 2 grace period to systems placed on the market before August 2, 2026. Anthropic's position is brief: the help center says developers should assess their own Article 50 obligations and promises technical guidance "as it becomes available." Treat the upstream mark as supporting evidence and document your own assessment. We covered a related line-drawing problem in the Act's one-third test for fine-tuners.
What a Detected Mark Proves and What It Misses
A detected mark tells you Claude was probably involved, and Anthropic says it does not settle more than that. The help center calls a detection "not fully conclusive," because Claude may have only proofread, translated or summarized the content, or it may have been edited afterward. A missing mark proves nothing either: older models, heavy paraphrasing, very short passages and unsupported platforms can all explain it.
The signal is also uneven by content type, which matters for how you describe it to customers. Anthropic's explainer says the mark is sparser on factual passages and on code, where exact wording leaves fewer choices, and that "a complete rewrite where every word is replaced will" remove it. It also says translations by Claude are marked, which localization teams should note. FIU's October piece featuring one of its AI researchers puts the practical floor at "about five sentences or more." The same piece says that if "only light proofreading is needed, it's unlikely the watermark will make its way into the text," a less cautious line than Anthropic's help center takes on proofreading.
Anthropic's explainer gives no numeric accuracy or false-positive rate, so you cannot yet put a figure in a risk register. OpenAI has published some for its own method: Search Engine Journal reports that in OpenAI's tests on 400-token English passages, swapping 25% of words for synonyms cut detection from about 92% to 17%. Those numbers do not transfer to Claude, though they do show how far detection of a word-choice mark can fall under heavy synonym substitution.
Access to detection is restricted, so you and your counterparties will not see the same evidence. The help center describes a private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations and EU civil society groups, plus "enterprises that need to check marks on their own Claude outputs" to meet their obligations under the Act. Your customer probably cannot test your deliverable, while a journalist might. Any contract that warrants "human-written" or "no AI-generated content" now has a checkable surface that one side controls, and a proofreading pass could complicate a dispute over it.
The objections have some weight. Users complained that proofreading or fact-checking with Claude could trigger detection and that bad actors can rephrase through another AI system, and some said output quality would drop, which Anthropic disputes. Business Insider found four Claude Max users who canceled, per Implicator, while Anthropic says it has seen no rise in cancellations. Neither side has published data that settles the quality question. Run your own evals on the models you pin if your output is sensitive to wording.
How Anthropic's Default Compares With OpenAI and Grok
Anthropic is the only one of these three vendors that documents a worldwide default with no opt-out. The policies differ enough that a multi-vendor routing layer will produce a mix of marked and unmarked text from the same product.
| Vendor | Default | Can you turn it off? | Detection access |
|---|---|---|---|
| Anthropic | On for supported models, worldwide | None described | Private preview, including enterprises checking their own outputs to meet AI Act obligations |
| OpenAI (textGrain) | On for eligible ChatGPT and Codex text in the EU, off in the API | API is opt-in for select models; EU ChatGPT opt-out not stated | Approved researchers and expert organizations, case by case |
| SpaceXSI (Grok API) | On from November 7, 2026 for listed models | Admin opt-out per model, unless locked | By application |
OpenAI's position is the strongest counterargument to Anthropic's. It told readers in 2024 that applying a watermark "to large volumes of text would lead to a large number of total false positives," and it says the EU-only rollout gives it room to learn from real use. For ChatGPT and Codex buyers outside the EU, that means no mark by default today. OpenAI's own post returned a 403 to our fetch, so these details come from Search Engine Journal and VKTR.
The Grok terms show what a vendor opt-out costs. SpaceXSI's documentation requires an admin to confirm "that an exception under applicable law applies to your use" and to accept being "solely responsible for any marking, disclosure, or other legal requirement." It does return an x-watermark response header, so you can log per-request status. Anthropic's documentation describes no equivalent, and the Sonnet 4.6 caveat means you cannot infer status for that model from the model name alone.
Some older Claude models are still unmarked. Anthropic's help center says it is "adding watermarks to outputs from models released before August 2, 2026," and Yahoo reports that models selectable back to Opus 3 are not yet covered. If you pin an old model for stability, as our deprecation comparison discusses, your EU exposure is a question to raise with Anthropic before December 2.
What to Do Before December 2
This Week:
- Search your code and framework configs for
temperature,top_pandtop_kon every Sonnet 4.6 and Opus 4.6 call, and list which workflows produce customer-facing or published text. - Ask your account team in writing whether Opus 4.6 is covered, from what date, and on which cloud platforms. The sources disagree, so get an answer on paper.
- Have your counsel pull customer contracts and agency agreements that warrant human authorship or the absence of AI-generated content, and flag each one that a Claude-assisted workflow touches.
By November 10:
- Note which Amazon Bedrock workloads run Mythos 5, Opus 4.7 or Sonnet 4.6, and record November 10 as the date Anthropic gives for rollout to finish. Anything you tell a regulator about earlier output should reflect that window.
- Submit the detector access request linked from the help center if you need to check your own outputs to meet AI Act obligations. The form registers interest in a private preview, so allow time for it.
Before December 2:
- Decide per pipeline, for low-temperature Sonnet 4.6 jobs, whether consistency or the mark wins. Then either drop the parameter and re-run your evals, or keep the request on
2023-06-01and record why. Add the decision to your audit trail. - Write down whether each EU-facing product makes you a provider, a deployer or both under Article 50, and add a human-visible disclosure wherever Article 50(4) applies. That duty sits outside the grace period, so it has applied since August 2. Our earlier Article 50 guide covers the chatbot side.
- Give your support and sales teams a short statement on the mark: it carries no customer identity, it can appear after light assistance, and neither its presence nor its absence proves authorship. Those three facts come straight from Anthropic's pages.
The Bottom Line
Statistical text watermarks have moved from research to vendor default in about four years: Anthropic traces its family of methods to a 2022 proposal by Scott Aaronson, Google introduced its text watermark in 2024, and the EU's Article 50 has now pushed Anthropic, OpenAI and SpaceXSI to ship their own. Compliance teams should expect the mark to become an assumed part of AI procurement, with the gaps (sampling settings, old models, short passages, restricted detectors) as the items that need answers in writing.
For most Claude customers, nothing breaks on November 3 and no request needs changing. The exposure sits in the pipelines that set sampling parameters and the contracts that promise human authorship. Start with the search for temperature, top_p and top_k in your Sonnet 4.6 calls.
Continue Reading
- EU AI Act's one-third test can make fine-tuners model providers
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- Adobe, Google and Getty Offer Different AI Image Indemnities
- Agent Audit Logging: Platform Logs Miss the Decision You Must Prove
- EU AI Transparency Law Starts Monday: Are You Ready?
- Anthropic's S-1 Shows $16.3B H1 Revenue and Q2 Profit