Claude’s text watermark reaches four older models on November 3. What it means

What changes on November 3

Anthropic emailed the organizations that use four older models on the Claude Platform, its API. The email, shared on X on November 1, names Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, and says: “On November 3, 2026, Anthropic will begin applying its text watermark to outputs from these four models, while rollout on Amazon Bedrock may take a few additional days. This is the same watermark our newer models, including Claude Opus 5.5, already carry.”

The newer models have carried it for a while. Anthropic announced the watermark on August 14, and its help center says “Claude models launched on or after August 2, 2026 support marking at launch.” Its release notes for September 1 say text from Claude Fable 5.1 and Mythos 5.1 “carries Anthropic’s text watermark,” and that “Marking requires no changes to your requests or response handling.”

The help center’s table, updated this week, lists 16 Claude models with text watermarks on Anthropic’s own products, and nine of them, Opus 4.8 and every 5-series model except Mythos 5, also on AWS, Google Cloud and Microsoft Foundry. The four older models in the email aren’t marked for the cloud partners yet. The table also lists Opus 4.5, Sonnet 4.5 and Haiku 4.5, which the email doesn’t mention, and we found no date for them.

The reason is the law. The EU AI Act has required providers to mark AI-generated content since August 2, 2026, and Anthropic signed the EU’s Code of Practice on the subject in July. The EU’s Digital Omnibus gives systems already on the market before August 2 until December 2, 2026 to comply, so older models have until then. Anthropic applies the watermark everywhere, not only in the EU: “We’re applying watermarking globally at launch because we don’t yet have a durable way to scope it by region.”

How the watermark works

When Claude picks the next word, it often has several that would do equally well, and a random number settles it. With watermarking, a key and the few words before settle it instead. The text reads the same, but someone with the key can check whether the choices fit the pattern. Anthropic’s method is a version of Google DeepMind’s SynthID-Text.

From Anthropic’s explainer:

• “Nothing is added to the text and there are no hidden characters.”
• “Watermarking doesn’t require extra tokens, and will not be more expensive.”
• “Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat.”
• “Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will.”
• “Detecting a watermark also doesn’t work well on small samples.”

It is thinner where there’s little choice. A factual sentence with one right answer has nothing to mark, and when Claude only fixes the grammar of your writing, nearly all the words are still yours. A translation is the opposite: “A translation produced by Claude carries a watermark,” because Claude chose every word.

Who can check for it

Not the public, for now. Anthropic’s detection API is 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),” and to companies that must verify watermarks for their own compliance. Organizations can register interest through a form linked from the help center. Files get a separate mark: supported files Claude makes, such as PNG and JPEG images, carry signed C2PA Content Credentials, which Anthropic’s free Claude Content Checker reads.

A result is a likelihood, not a verdict. Anthropic’s help center says “A detected mark provides a signal that content was processed by Claude, but is not fully conclusive,” because Claude may only have proofread, translated or summarized it. And a missing mark proves nothing either: the text may be short, heavily edited, or from an older model. Anthropic’s explainer adds that it “cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’”

What it means for coding agents

The watermark comes from the model, not the app. Anthropic’s help center says it “will be present no matter which Claude product or surface the text comes from,” so it is in Claude Code’s output as much as Claude’s chat. The same is true when Claude Code runs on your own Claude plan in another app, such as Enjoy.

Code itself carries little of it. Anthropic says “code—which in very many cases has to be exact—has generally less watermarking than some other forms of text.” Comments and other free choices can carry it, but it has “a negligible effect on the actual code produced.”

If you call Opus 4.6 or Sonnet 4.6 through the API, one detail matters. The Decoder reports that Anthropic’s email says the watermark applies to those two only when temperature, top_p and top_k stay at their defaults, and quotes it: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” Anthropic’s public pages don’t mention this. Claude 4.7 and later models accept only the default values anyway, as our memo on API version 2026-10-21 explains.

How ChatGPT, Grok and Gemini compare

Anthropic isn’t alone. OpenAI signed the same EU code, SpaceXSI cites the EU AI Act too, and each lab has made different choices:

	Claude	ChatGPT and OpenAI’s API	Grok API	Gemini
What’s marked	Text from supported models, in every Claude product and the API, worldwide	ChatGPT text in the EU; API text if you opt in	Text, structured outputs and tool-call arguments from seven models	Text from the Gemini app and web
Since	Newer models already; four older ones from November 3	Announced October 5, 2026	November 7	Not dated on Google’s page
Can you turn it off?	No setting described	API: off unless you opt in; ChatGPT: not said	Team admins can opt out per model	Google’s page doesn’t say
Who can check text	Approved organizations	Approved researchers and organizations	By application	Google’s page doesn’t say

OpenAI. Its announcement of October 5 said it would add its textGrain watermark to “eligible ChatGPT and Codex users across all plans in the EU only,” and “We are not making text watermarking a global default at launch.” Its help center now says that in the EU, ChatGPT-generated text includes it, and that API customers anywhere can turn it on in project or organization settings for select models. OpenAI is plain about the limits: replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

SpaceXSI, the company behind Grok, says in its API docs that “Starting November 7, 2026, responses from these models include a watermark,” naming grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 models, on every API endpoint. A response header, x-watermark: true, tells you a response was marked. Team admins can opt out per model, but opting out “asks you to confirm that an exception under applicable law applies to your use,” and some models “may be locked on.” The docs are about the API and don’t mention the Grok apps or Grok Build.

Google says on its SynthID page: “We’ve expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience.” The page doesn’t say whether the Gemini API or Antigravity marks text.

What to do

• If you write with Claude: nothing changes in what you see, and nothing identifies you. If it matters that a client, school or publisher knows Claude helped, say so. A watermark is only a signal, and its absence proves nothing either way.
• If you build on the Claude API: you don’t need to change any code. If you rely on the watermark for your own EU compliance, keep the sampling settings at their defaults on Opus 4.6 and Sonnet 4.6, and remember Anthropic’s advice: “If you deploy Claude in your own product, you should independently assess what Article 50 requires of your products and services.”
• If you use Grok’s API: expect the watermark from November 7 unless your team admin opts out, and check the x-watermark header to see what you got.
• If you need to detect AI text: apply to each lab separately. Each lab’s watermark has its own key, and Anthropic says its check can’t tell whether text came from a different AI.

Anthropic hasn’t said when Opus 4.5, Sonnet 4.5 and Haiku 4.5 get the watermark, or when the four older models get it on Google Cloud and Microsoft Foundry. We’ll update this memo when it does.