top of page

Claude text watermark extends to Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5

2 minutes ago
7 min read
Cover showing the Claude logo and the title about the Claude text watermark extending to Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. Data Studios original cover

Anthropic has added its text watermark to the four older Claude models that were still outside it. An entry dated November 16, 2026 in the Claude Platform release notes states that text generated by Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5 now carries Anthropic's text watermark on supported platforms, with the list of those platforms kept in the company's support article on how Claude marks AI-generated content. When Anthropic extended the watermark to Claude Mythos 5, Opus 4.7 and Sonnet 4.6 on November 3, these four models were the ones its documentation still described as producing text with no watermark.


The new entry carries the same technical condition that Anthropic attached to Sonnet 4.6 two weeks ago. If a request to any of the four models sets `temperature`, `top_p` or `top_k` to a non-default value, the generated text might not carry the watermark. The Claude Platform page on the text watermark now applies that guidance to six models in total: Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 and Claude Opus 3, which Anthropic offers only on request.


With the November 16 change, every Claude model in the support article's marking table produces watermarked text, from Claude Fable 5.2 and Mythos 5.2 down to Haiku 4.5. The support article itself had not caught up on the day of the release note: it still listed Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 as models that attach Content Credentials to generated files without marking their text, so the release notes are the authoritative record of the change for now.


··········


WHICH CLAUDE MODELS CARRY THE TEXT WATERMARK AFTER THE NOVEMBER 16 UPDATE.


Anthropic's marking system has two layers. The text watermark is applied at the model level and sits in the word choices of the output, while Content Credentials based on the C2PA standard are signed metadata attached to files that Claude creates, such as PNG and JPEG images. The November 16 entry changes only the first layer, since the four models already attached Content Credentials to supported files. The table below combines the support article's model list with the November 3 and November 16 release notes and the sampling guidance on the text watermark page.


........


Model

Text watermark before Nov 16

Text watermark from Nov 16

Sampling caveat

Claude Fable 5.2, 5.1, 5

Yes

Yes

No

Claude Mythos 5.2, 5.1, 5

Yes

Yes

No

Claude Opus 5.5, 5, 4.8, 4.7

Yes

Yes

No

Claude Sonnet 5.5, 5 and Haiku 5.5

Yes

Yes

No

Claude Sonnet 4.6

Yes, since Nov 3

Yes

Yes

Claude Opus 4.6

No

Yes

Yes

Claude Opus 4.5

No

Yes

Yes

Claude Sonnet 4.5

No

Yes

Yes

Claude Haiku 4.5

No

Yes

Yes


........


The sampling caveat divides the lineup into two groups. On the newer models, which were built with marking from launch or received it earlier, Anthropic's documentation sets no condition on sampling parameters. On the five older models in the lower part of the table, the watermark is reliable only when the request leaves `temperature`, `top_p` and `top_k` at their defaults. Anthropic has not explained the technical reason for the difference, and the release note does not say whether the condition will be lifted later.


The support article lists the surfaces where text watermarks apply: the Claude Platform API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, together with supported models accessed through AWS, Google Cloud and Microsoft Foundry. Anthropic says the marking applies worldwide, while some platforms or features may not support every type of mark. The support article also says that the Amazon Bedrock rollout for certain models was to be completed by November 10, while the November 16 entry gives no separate schedule for the four models on the cloud partners.


··········


CLAUDE API SAMPLING PARAMETERS AND THE WATERMARK ON OLDER MODELS.


For developers who call the four models through the Messages API, the change needs no code update in most cases. Anthropic's text watermark page says the watermark does not change the meaning, quality or readability of the output, adds no tokens or hidden characters, carries no information about the user or the organization, and requires no change to requests or response handling. Billing is unaffected, because watermarking adds no tokens to the request or the response.


The exception concerns applications that tune sampling. Many production integrations set `temperature` to 0 for extraction or classification tasks, or raise it for creative text, and older frameworks often send `top_p` as well. From November 16, those requests to Opus 4.6, Opus 4.5, Sonnet 4.5 or Haiku 4.5 may return text without the watermark. Anthropic's documentation tells teams that need the watermark to omit all three parameters, and it warns that a framework or client library may set them without the developer noticing, so the request body that actually reaches the API is what counts. The documentation gives this minimal request as the pattern:


import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",  # check the model ID in the docs
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize the water cycle."}],
    # Do not set temperature, top_p, or top_k
)
print(response.content[0].text)

Sampling settings were already a source of friction on newer models, several of which reject a custom `temperature` outright, a problem covered in our explainer on the Unsupported parameter: temperature error in the OpenAI, Claude and Gemini APIs. The older models still accept the parameters, so a request that sets them will succeed and return normal text, and the only difference is that the watermark may be absent. Nothing in the API response signals whether a given output carries the mark.


··········


WHAT THE NOVEMBER 16 CHANGE MEANS FOR CLAUDE APP, API AND CLOUD USERS.


The effect depends on the model in use and on how requests are configured. The following table sums up the situations described in Anthropic's documentation.


........


User

Model in use

Effect from November 16

Action

Claude app, Claude Code or Cowork user

Any listed model

New text carries the watermark

None

API developer, default sampling

Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5

New text carries the watermark

None

API developer, custom sampling

Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5

Watermark may be missing

Omit temperature, top_p and top_k if the mark is needed

API developer, custom sampling

Sonnet 4.6

Watermark may be missing, as since Nov 3

Same as above

API developer, any sampling

Opus 4.7 and newer models

New text carries the watermark

None

AWS, Google Cloud or Microsoft Foundry customer

Supported listed models

Watermark applies on the partner platform

Check the support article for the platform


........


Two of the four models are close to the end of their life on Anthropic's own platforms. Anthropic's model deprecations page lists Claude Sonnet 4.5 (`claude-sonnet-4-5-20250929`) as deprecated since September 30, 2026, with a tentative retirement date of November 30, 2026 and Claude Sonnet 5.5 as the recommended replacement. Claude Opus 4.5 is still active, with a retirement date of not sooner than November 24, 2026, while Claude Opus 4.6 is active until not sooner than February 5, 2027 and Claude Haiku 4.5 remains active with no retirement date set. Those dates apply to the Claude API, Claude Platform on AWS, Anthropic-managed Claude models on Google Cloud and Microsoft Foundry, while Amazon Bedrock and Google-managed deployments set their own schedules.


For teams still pinned to these model IDs, the update means that text produced in the last weeks of Sonnet 4.5 and in the remaining life of Opus 4.6 and Haiku 4.5 can be attributed to Claude by a detector holding Anthropic's key, provided the requests use default sampling. Our guide to Claude, Gemini and DeepSeek API model names, aliases, pinned versions and retired IDs explains which identifiers point to which model versions, which helps when checking what a production system actually calls.


··········


HOW CLAUDE'S TEXT WATERMARK WORKS AND THE EU DEADLINE BEHIND IT.


The mechanism has not changed since Anthropic described it in a post on August 14, 2026, which our report on the November 3 extension of the Claude text watermark to Mythos 5, Opus 4.7 and Sonnet 4.6 covered in detail. When Claude selects each word from a set of likely candidates, the watermark replaces the usual random choice with one derived from a secret key and the preceding words, leaving a statistical pattern that a detector with the key can measure. Anthropic says the method is a version of SynthID-Text, the technique Google DeepMind published in 2024, that detection gives a probability rather than a verdict, and that confidence grows with the length of the passage.


The design explains the sampling caveat in general terms, since the watermark lives in the way words are sampled, although Anthropic has not said why only the older models are sensitive to custom parameters. It also explains the limits Anthropic lists in the support article: very short passages, heavily edited or paraphrased text and output where the wording is largely fixed may carry too little signal to detect, and a missing mark does not prove that a text was written without AI.


Detection of the text watermark is in private preview. The support article lists regulators, law enforcement, media organizations, fact-checkers, independent researchers, educational organizations, EU civil society groups and enterprises that must verify their own Claude outputs as eligible applicants, through a Claude Watermark Detector access request form, while Content Credentials on files can be checked with the free Claude Content Checker. Anthropic ties the rollout to the EU AI Act: models launched in the EU on or after August 2, 2026 support marking at launch, and older models fall under a transition period during which Anthropic has been adding support model by model. The support article also tells developers building on Claude to assess their own transparency obligations under Article 50, because the watermark marks the model's output and does not cover every disclosure duty that a deployer has toward its users. xAI took a similar step on its developer platform, as described in our article on the Grok API invisible text watermark enabled by default from November 7.


··········


CLAUDE TEXT WATERMARK ON OPUS 4.6, OPUS 4.5, SONNET 4.5 AND HAIKU 4.5: CONFIRMED FACTS AND OPEN QUESTIONS.


The November 16 release note confirms that Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 now produce watermarked text on supported platforms, which completes the coverage of the models in Anthropic's marking table. It also confirms that on these four models, as on Sonnet 4.6, a non-default `temperature`, `top_p` or `top_k` can leave the output unmarked, and Anthropic's documentation gives omitting those parameters as the only control available to developers who need the mark.


Several details are still missing. The support article had not been updated to reflect the change on the day of the announcement, the release note gives no separate timing for the cloud partner platforms, and the text watermark page lists Claude Opus 3 among the models covered by the sampling guidance without further explanation in the release notes. Anthropic has not published detection accuracy, false-positive rates or a minimum passage length, and the detection API has no general availability date. For a company that needs Claude-generated text to remain identifiable, the configuration that Anthropic's documentation supports today is any model in the table above and, on the five older models, requests that leave all three sampling parameters at their defaults.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page