Claude text watermark now covers Mythos 5, Opus 4.7 and Sonnet 4.6

19 minutes ago

7 min read

Anthropic has extended its text watermark to three more Claude models. According to an entry dated November 3, 2026 in the Claude Platform release notes, text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on the platforms listed in the company's support article on how Claude marks AI-generated content. Until this change, Anthropic's documentation placed Mythos 5 among the models that attach Content Credentials to generated files but put no watermark in their text, and listed Opus 4.7 and Sonnet 4.6 as models not yet supported for text watermarking.

The update comes with one technical caveat that developers need to know. On Claude Sonnet 4.6, Anthropic says text generated for a request that sets `temperature`, `top_p` or `top_k` to a non-default value might not carry the watermark. The other two models in the update have no such note in the release entry. Anthropic's text watermark adds no tokens and no hidden characters to the output, does not change the price of a request, and carries no information that identifies the user, the organization or the conversation.

With this step, the watermark reaches models that were launched before August 2, 2026, the date on which the European Union's requirement for marking AI-generated content took effect. Anthropic had said in August that the EU law includes a transition period for models launched before that date and that it would add watermarking to those older models over the following months. Four older models remain outside the text watermark for now: Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5.

··········

CLAUDE TEXT WATERMARK: WHICH MODELS CARRY IT AFTER THE NOVEMBER 3 UPDATE.

Anthropic's support article sorts Claude models into three groups: models whose text carries the watermark on Anthropic's own surfaces and on cloud partners, models that only attach a C2PA Content Credential to the files they generate, and models not yet supported for text watermarks. The November 3 release note moves three models into the first group. The table below combines the support article's list with the change announced in the release notes.

........

Model	Text watermark before Nov 3	Text watermark after Nov 3	File Content Credentials (C2PA)
Claude Fable 5.1, Fable 5	Yes	Yes	Yes
Claude Mythos 5.1	Yes	Yes	Yes
Claude Opus 5.5, Opus 5, Opus 4.8	Yes	Yes	Yes
Claude Sonnet 5.5, Sonnet 5, Haiku 5.5	Yes	Yes	Yes
Claude Mythos 5	No (C2PA only)	Yes	Yes
Claude Opus 4.7	No	Yes	Yes
Claude Sonnet 4.6	No	Yes, with a sampling caveat	Yes
Claude Opus 4.6, Opus 4.5	No	No	Yes
Claude Sonnet 4.5, Haiku 4.5	No	No	Yes

........

The support article says the marks cover the Claude Platform API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, and that supported models accessed through AWS, Google Cloud or Microsoft Foundry carry the watermark as well. Because the watermark is applied at the model level, Anthropic says it appears regardless of the surface on which the model is used, and the marking applies wherever Claude is offered, in the EU and elsewhere.

The practical consequence is that the model choice now decides whether a piece of text can later be attributed to Claude. A company that still runs Opus 4.7 or Sonnet 4.6 in production because of cost, tested prompts or a pinned model ID will, from this update, produce watermarked text without changing a line of code. A team that stays on Opus 4.6 or Sonnet 4.5 produces text that, for now, carries no text watermark at all.

··········

HOW ANTHROPIC'S CLAUDE TEXT WATERMARK WORKS.

Anthropic described the mechanism in a post published on August 14, 2026. When Claude writes, it picks each word from a set of likely candidates, normally using a random number source to choose among them. The watermark replaces that random source with one derived from a secret key and the preceding words, so the word choices follow a statistical pattern that a detector holding the key can recognize. Anthropic says the method is a version of SynthID-Text, the technique Google DeepMind published in Nature in 2024.

Detection does not return a yes-or-no verdict. It checks how closely a word sequence matches what Claude would produce with the key and assigns a probability that Claude was involved in writing it. Anthropic says confidence increases with the length of the passage and that detection works poorly on short samples. A positive result indicates that Claude was likely involved; it cannot confirm human authorship and cannot identify text from other AI systems.

The company lists several limits that follow from the design. The watermark only exists in words Claude chooses, so it is sparse or absent where the output is fixed, such as a factual answer or most code, although it can still appear in free-form parts like code comments. Light proofreading of human text may leave too few watermarked words to detect, while translations are watermarked because Claude picks every word. Anthropic says light edits will probably not remove the watermark and a full rewrite will.

Files follow a separate path. When Claude produces a supported file type such as PNG, JPEG or SVG, it attaches a signed Content Credential that follows the C2PA open standard as metadata, and the file content itself is not altered. Anthropic offers a free Claude Content Checker for those credentials, while text watermark detection runs through an API in private preview for eligible organizations, including regulators, law enforcement, media organizations, fact-checkers, independent researchers, educational organizations, EU civil society groups and enterprises with their own verification obligations.

··········

WHAT CHANGES FOR CLAUDE API DEVELOPERS USING SONNET 4.6.

For most users the update requires no action. Anthropic's documentation states that the watermark does not change the meaning, quality or readability of the output, adds no tokens, and does not require any change to requests or response handling. Billing is unaffected, since the August announcement says watermarking requires no extra tokens and will not be more expensive.

The exception concerns sampling parameters on Sonnet 4.6. The text watermark page in the Claude Platform documentation says that on this model, a request that sets `temperature`, `top_p` or `top_k` to a non-default value might produce text without the watermark, and it advises developers who rely on the watermark to omit all three parameters, including any that a framework or client library sets on their behalf. Anthropic's documentation gives this minimal Messages API call as the pattern to follow:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="<your-model-id>",  # see the support article for supported models
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize this report."}],
    # No temperature, top_p, or top_k parameters
)
print(response.content[0].text)

The situations in which the change has a concrete effect can be summarized by type of user.

........

User	Model in use	Effect of the November 3 update	Action
Claude app, Claude Code or Cowork user	Opus 4.7 or Sonnet 4.6 selected	New text carries the watermark	None required
API developer	Opus 4.7 or Mythos 5	New text carries the watermark	None required
API developer	Sonnet 4.6 with default sampling	New text carries the watermark	None required
API developer	Sonnet 4.6 with custom temperature, top_p or top_k	Watermark may be missing	Omit all three parameters if the watermark is needed
API developer	Opus 4.6, Opus 4.5, Sonnet 4.5 or Haiku 4.5	No text watermark yet	Plan for watermarking to arrive later

........

Developers building products on Claude should also keep in mind that Anthropic's support article tells them to assess their own transparency obligations; the watermark marks the model's output, and it does not by itself satisfy every disclosure duty a deployer may have toward its users. For the full list of current and older model IDs, our guide to Claude, Gemini and DeepSeek API model names, aliases and retired IDs covers which identifiers point to which models, and our guide to the Claude models available across the app and the API explains how access differs by plan.

··········

THE EU TRANSPARENCY RULES BEHIND CLAUDE'S WATERMARK ROLLOUT.

Anthropic has tied the watermark to the EU AI Act. The support article states that new Claude models launched in the EU on or after August 2, 2026 support marking at launch, and that the law includes a transition period for systems launched before that date. In July 2026, Anthropic was one of about 190 signatories of the EU Code of Practice on Transparency of AI-Generated Content, according to its August post.

Other model providers are moving in the same direction on their developer platforms. xAI said its Grok API will add invisible watermarks to text from seven models by default from November 7, as reported in our article on the Grok API text watermark. Anthropic is also working on provenance from a different angle: its open-source FlashProve project, covered in our report on proving which AI model wrote a reply, addresses the question of which model served a given response rather than whether a text came from Claude.

Anthropic's watermark carries no identifying information and cannot be traced to a specific person, organization or chat, according to the company, and it does not change ownership, authorship or user rights under Anthropic's terms. The design therefore answers a narrow question, whether Claude was likely involved in producing a text, and leaves attribution to individuals outside its scope.

··········

CLAUDE TEXT WATERMARK COVERAGE: CONFIRMED FACTS AND OPEN QUESTIONS.

The release note establishes three facts: Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now produce watermarked text on the platforms Anthropic lists, Sonnet 4.6 may skip the watermark when developers change its sampling parameters, and the mechanism is the same one Anthropic has applied to its newer models since August. The support article confirms that watermarking is applied at the model level across Anthropic's own surfaces and the AWS, Google Cloud and Microsoft Foundry platforms.

Several points remain undisclosed. Anthropic has not given a date for adding the text watermark to Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, beyond its earlier statement that the older models would be covered over the coming months. It has not published detection accuracy or false-positive rates, nor a minimum text length for reliable detection, and the detection API remains in private preview with no general availability date. For teams that depend on Claude-generated text being identifiable, the safe configuration today is a model from the watermarked list and, on Sonnet 4.6, requests that leave temperature, top_p and top_k at their defaults.

··········

·····

DATA STUDIOS

·····

datastudios.org