Anthropic Watermarks 4 More Claude Models: The 1 Setting That Removes the Mark on 2 of Them

On October 31, 2026, Anthropic extended its hidden text watermark to four more Claude models, and in the same message it disclosed a limit that matters more than the expansion itself. The four models are Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Anthropic told API customers about the change by email, and newer models such as Claude Opus 5.5 already carry the mark. The mark itself is an imperceptible pattern woven into word choice, so generated text reads normally while carrying a statistical signal that can be checked later. The catch is on two of the four models: on Opus 4.6 and Sonnet 4.6, the watermark only applies when three sampling settings stay at their defaults. Change any one of them, and the output is unmarked. Anthropic's own advice is direct: if you rely on the watermark for your own compliance, keep those settings at their defaults.

That advice is the reader action in this article. If you use Claude through the API, check your sampling settings before you claim that your output is watermarked. If you buy or review content that claims to be watermarked Claude output, do not treat the absence of a mark as proof that a human wrote it. And if you run a product that depends on detecting AI text, do not build a rule that assumes every Claude output from these models will carry a mark, because on two of them it will not, whenever a developer has tuned the randomness.

What is the Claude watermark, and what does it not do?

Anthropic describes the watermark as a pattern in word choice, built with a variant of Google DeepMind's SynthID text method, which tweaks the randomness behind word selection to produce a statistically detectable pattern. Anthropic says the mark does not change the meaning, quality, or readability of the response, does not add extra text or tokens, does not affect pricing or the API format, and does not include information about the user or the conversation. The mark travels with the text when it is copied and pasted, and Anthropic says it may survive some editing. Customers cannot turn it off directly, which is the point of a transparency mark: it is embedded at the model level, so it works across every platform that hosts these models, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Amazon Bedrock may receive the change a few days later than other platforms.

The mark is not a visible label, not a badge on the page, and not a guarantee about who wrote a document. Anthropic's help page is clear on the limit that matters most for anyone checking text: a detected mark means the text may have come from Claude, not that Claude definitely wrote it, because people also use the model to proofread or translate human writing. A text that was edited by Claude, or translated by Claude, can carry a mark even though a human wrote the original draft. The reverse limit is just as important, and it is the subject of this article: on Opus 4.6 and Sonnet 4.6, unmarked text does not prove a human wrote it, because a single changed sampling setting is enough to avoid the mark. A watermark system that can be absent for a legitimate technical reason cannot be used as a simple human or AI test, and any workflow that treats it that way will produce false conclusions in both directions.

Detection is also limited in practice. Anthropic operates a watermark detection API in private preview, available to eligible organisations such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, and EU civil society groups, plus firms that must verify watermarking for their own compliance. There is no public text checker for ordinary readers, so a reader cannot paste a paragraph into a public tool and get a reliable Claude watermark verdict. For files, Anthropic offers a free Claude Content Checker that reads C2PA Content Credentials, but that is a different check from the text watermark, and it applies to files that carry those credentials. If someone offers you a public watermark checker for Claude text, treat it with caution, because Anthropic's own text detection is not generally available.

Which models are covered, and where is the loophole?

The October 31 expansion covers four models, and the settings limit covers two of them. Mythos 5: watermark added, no such caveat mentioned in Anthropic's email, which suggests the mark applies regardless of sampling settings, but the email does not explicitly confirm that. Opus 4.7: watermark added, same note as Mythos 5. Opus 4.6: watermark added, mark only applies when temperature, top_p, and top_k are all at their defaults. Sonnet 4.6: watermark added, same limit as Opus 4.6. Opus 5.5 and other newer models: already watermarked before this expansion, no settings loophole reported for these newer models.

Temperature, top_p, and top_k are the three settings that control how much randomness the model uses when picking words. Developers change them routinely, for good reasons: a lower temperature for more predictable, factual output, a higher temperature for more varied creative output, and top_p and top_k adjustments to narrow or widen the pool of candidate words. None of those changes is suspicious on its own. They are normal tuning choices in almost every API application. That normality is what makes the loophole awkward for compliance. A developer does not need to intend to evade a watermark to produce unmarked output on Opus 4.6 or Sonnet 4.6. They only need to have tuned one setting, as most production applications do, and the mark disappears without any warning in the output itself.

Anthropic warns customers who rely on the watermark for compliance to keep these settings at their defaults. That warning is useful, but it is also a condition that many real applications cannot meet without changing their output quality. A factual support bot, a code assistant, and a creative writing tool will often use different sampling settings for good product reasons. If such an application also needs watermarked output for compliance, it now has a direct conflict on Opus 4.6 and Sonnet 4.6: keep defaults and accept the output style that comes with them, or tune the settings and lose the mark. The honest resolution is to choose the newer models where the mark does not carry this caveat, or to treat the watermark as one signal among others, not as the compliance mechanism itself. What a team should not do is keep tuned settings, keep claiming watermark coverage, and discover the gap only when a regulator or a customer asks for proof.

Why is Anthropic doing this now?

The driver is the EU AI Act's December deadline, and the staged rollout that comes with it. Every model released after August 2, 2026 shipped with the watermark from day one, while older models are being brought in gradually under the law's transition period. Anthropic signed a European code of practice for labelling AI output, tied to the AI Act, and the watermark is part of meeting that transparency commitment. The October 31 expansion is therefore not a surprise feature, it is a scheduled retrofit of older models before a deadline, and the settings limit on two older models looks like a technical constraint of retrofitting a mark onto models that were not built with it from the start, rather than a deliberate back door. That explanation does not remove the compliance risk, but it does explain why the limit exists on older models and not on newer ones.

The wider context also matters, because Claude is not the only system moving in this direction. OpenAI launched its own watermarking system, textGrain, in October 2026, but it only activates by default for ChatGPT and Codex users in the EU, as reported by The Decoder. That EU first pattern is the same pattern as Anthropic's: transparency rules with EU deadlines are driving the first wave of watermarking, and other regions get the same technology through the same models, but not always with the same defaults. A business operating in the US, Canada, or the UK should not assume that an EU driven watermark behaves the same way in every product and every region. Check the model, the platform, and the settings, not just the brand name.

There is also a quality debate that a careful reader should know about, even though Anthropic says the mark does not change quality. Blogger John Gruber has argued that choosing words partly to create a watermark can reduce precision, because the model is no longer always picking the best word, it is picking a word that also fits the pattern. Anthropic's position is that the mark does not change meaning, quality, or readability. Both claims can be partly true at different levels: a mark can be imperceptible to a casual reader and still change a word choice in a technical document where the exact word matters. For most everyday writing, the difference will not be visible. For legal, medical, or technical text where precision is the product, test watermarked output on your own documents before you rely on it, and do not assume that imperceptible means identical.

What should you do this week?

If you build with Claude through the API, do three checks this week, and write down the answers. First, list every model you call in production, and mark which ones are in the October 31 expansion, which ones were already watermarked, and which ones are not yet watermarked at all. Anthropic's help page table, as re-read on October 31, still showed no text watermark yet on some older models, including Opus 4.5, Sonnet 4.5, and Haiku 4.5, so do not assume that all Claude models are covered because four more were added. Second, list the sampling settings for each application that calls Opus 4.6 or Sonnet 4.6, and flag any application where temperature, top_p, or top_k is not at the default. Those applications are producing unmarked output today, even if your documentation says otherwise. Third, decide, in writing, whether your watermark reliance is a compliance requirement, a nice to have signal, or not a factor at all, and change either the model or the settings to match that decision. A decision that says we rely on the watermark but we also tune all three settings is not a decision, it is a contradiction.

If you review, publish, or buy content, change one habit: stop treating unmarked text as human written, and stop treating marked text as proof of full AI authorship. A marked text may have come from Claude, but it may also be human text that Claude proofread or translated. An unmarked text may be human, or it may be Claude output from a tuned Opus 4.6 or Sonnet 4.6 call, or from a model that is not yet watermarked, or from another provider entirely. The watermark is a useful signal for platforms and regulators who have detection access, and it is a weak signal for everyone else. Use it as one input, alongside source records, author history, and the actual quality of the work, not as a verdict.

If you are a publisher or a business owner, this story connects to yesterday's robots.txt story, but it is a different control. The robots.txt token controls whether your content may be used for training. The Claude watermark controls whether text that Claude generates can later be identified as Claude generated. One protects your input, the other labels an output. You need both ideas in your policy, but you cannot substitute one for the other, and neither one, on its own, tells you whether a document in front of you was written by a person or by a model.

Key takeaways: Anthropic is extending its invisible text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6, as told to API customers on October 31, 2026, with newer models such as Opus 5.5 already marked. On Opus 4.6 and Sonnet 4.6, the mark only applies when temperature, top_p, and top_k stay at their defaults. Changing any one of them is enough to produce unmarked output. Anthropic warns customers who rely on the watermark for compliance to keep those settings at their defaults, which conflicts with normal tuning in many production applications. The email does not mention the same caveat for Mythos 5 or Opus 4.7, which suggests those two are marked regardless of settings, but that is not explicitly confirmed. Detection is in private preview for eligible organisations only. There is no public text checker, so ordinary readers cannot verify a mark themselves. A detected mark means text may have come from Claude, not that Claude wrote it fully, because proofreading and translation can also carry a mark. Unmarked text does not prove human authorship. The rollout is driven by the EU AI Act December deadline, with models released after August 2, 2026 marked from day one and older models retrofitted in stages.

Related links: Top Use AI Blog, Is AI voice cloning legal, Best free SEO tools

Affiliate disclosure: No affiliate link is used in this article. Links to Anthropic, The Decoder, and TopUseAI posts earn no commission.

Sources: Anthropic email to API customers, October 31, 2026, as reported by The Decoder, October 31, 2026; Anthropic help page on how Claude marks AI-generated content, as re-read October 31, 2026; Anthropic EU AI Act Code of Practice transparency commitments, models released after August 2, 2026 marked from launch.