Skip to content

Does setting temperature remove Claude's text watermark?

Unity Horizon9 min read

On six Claude models Anthropic says a non-default temperature, top_p or top_k may strip the mark. Two of eight client libraries send one anyway.

On six of them, yes. Anthropic tells you to omit temperature, top_p and top_k from a request if you need the output to carry its text watermark, and six of the nineteen Claude models on the coverage list carry that condition. On the other thirteen the question cannot come up, because the API returns a 400 for all three.

Checked by Unity Horizon on 20 November 2026.

Which Claude models can lose the watermark?

The text watermark page is short and one line of it is the whole problem. "If you need your output to carry the watermark, omit temperature, top_p, and top_k from your requests. If a request sets any of them to a non-default value, the output might not carry the watermark." The default for temperature is 1.0.

The coverage page names nineteen models, on Anthropic's own surfaces and at its three cloud partners. Six of the nineteen carry a dagger. The footnote says their "outputs may not reliably carry the text watermark when temperature, top_p, or top_k is set to a non-default value."

Model Accepts a non-default sampling value Anthropic warns the mark may be missing
Opus 4.7 and later, plus the Fable and Mythos families, thirteen models no, 400 error no
Opus 4.6 yes yes
Opus 4.5 yes yes
Sonnet 4.6 yes yes
Sonnet 4.5 yes yes
Haiku 4.5 yes yes
Opus 3, access on request yes yes

The line between the two halves of that table is one model generation. Anthropic's parameter deprecation note puts it plainly: the three parameters are removed in API version 2026-10-21 and return a 400 error, and on earlier API versions they already return one "when set to a non-default value on Claude 4.7 and later models". Below that line the value is accepted, and nothing in the response tells you the mark has gone.

When did the older models get a watermark at all?

This month. The API release notes record text watermarking reaching Mythos 5, Opus 4.7 and Sonnet 4.6 on 3 November 2026, and Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 on 16 November 2026, four days before we checked. Anthropic's announcement of 14 August 2026 had promised the retrofit over "the coming months". It arrived sixteen days before the 2 December date the EU set for marking.

Claude Sonnet 4.5 was deprecated on 30 September 2026 and retires on 30 November. It was given a watermark fourteen days before it stops answering requests.

Why would a sampling parameter touch a watermark?

Because the mark is made out of the choices the model had. Anthropic calls its method "a version of the SynthID-Text approach" and says it "only changes the source of the randomness used to pick among words". The open-source implementation of the same method describes the step in the same place, "subtly manipulating the probabilities of token selection during text generation in a manner that can be reliably recovered later for verification".

So lower the temperature and the leading token wins more often. Set top_k and the candidates the pattern would have used are gone before the pattern runs. Anthropic says as much about where its mark goes thin, without connecting any of it to a parameter: the watermark "is sparser on factual passages", code carries "generally less watermarking than some other forms of text", and detection "doesn't work well on small samples".

The mark lives where the model had a choice. Take the choice away and there is nothing left to mark.

Does your client library send a temperature you never set?

Two of the eight we read do. On 20 November 2026 we read the request-building code of eight widely used clients on their default branches, with one question each: when the caller sets no sampling parameter, does the body sent to the Messages API contain temperature, top_p or top_k?

Client With no sampling parameter set by the caller
langchain-anthropic nothing, the payload is filtered for nulls before the call
Vercel AI SDK nothing, only what the caller passes
pydantic-ai nothing, sampling settings ride in extra_body only when set
LiteLLM nothing, a parameter is mapped only when supplied
Agno nothing, all three default to None and are sent only if set
CrewAI nothing, None passes through to LiteLLM
llama-index-llms-anthropic temperature 0.1, on every request to a model not on its strip list
Flowise ChatAnthropic temperature 0.9, pre-filled in the node's form

LlamaIndex reads DEFAULT_TEMPERATURE from its core constants, which is 0.1, and puts it in the request. It also carries a tuple named ANTHROPIC_NO_TEMP_MODELS and deletes the parameter for everything on it: Opus 4.7, Opus 4.8, Opus 5, Opus 5.5, Fable 5, Sonnet 5, Sonnet 5.5 and Haiku 5.5. Those are the models that answer with a 400. Sonnet 4.6, Sonnet 4.5, Opus 4.6, Opus 4.5 and Haiku 4.5 are absent from the tuple, and those are five of the six models with the footnote.

The list removes the parameter where the API complains. It keeps it where only the watermark pays.

Flowise fails differently. Its ChatAnthropic node renders a Temperature field with 0.9 already sitting in it and sends that value whenever the chosen model accepts sampling, so the number reaching the API is the one the node shipped with rather than one anybody picked.

Can you check whether your own text carries the mark?

Not at any of the three vendors, unless one of them approves you first. Anthropic's coverage page says "Watermark detection is currently in private preview", open to regulators, law enforcement, media and researchers, and to "select enterprises checking their own outputs". The free Content Checker it offers reads Content Credentials on files, which means a C2PA manifest wrapped around a PNG. A sentence has no manifest to wrap. OpenAI, announcing its own text watermark on 5 October 2026, said "detector access will initially be limited to approved researchers and expert organizations".

So a provider is told to mark its output, and told that one request field can remove the mark. No vendor will let that provider check.

Vendor Text watermark on API output Default state What the vendor says weakens it Who may detect
Anthropic nineteen models, worldwide on non-default temperature, top_p or top_k on six models; a complete rewrite; factual prose and code private preview
OpenAI select models, per project or organization off short passages and later edits approved researchers and expert organizations
Google, Gemini API not documented not stated not stated no public tool for text

Does the Gemini API watermark text?

Google has not written it down, and the question keeps being asked. On the official Gemini API forum a developer asked on 5 August 2026 whether API text carries SynthID or any machine-readable provenance signal, and asked for documentation they could cite. The answer came the same day: "Generated text from the API is NOT SynthID-watermarked", and native text watermarking "is not planned at the moment". On 19 August the same account retracted it after checking with the team, and wrote that "text generated via the Gemini API IS actually SynthID-watermarked".

Four more people asked in the following three months, about Vertex AI in an EU region, about specific Flash models, about whether JSON output changes anything, and about which tool a third party could verify with. One of them runs a WhatsApp assistant for restaurants in Spain. A reply on 16 November 2026 said Google applies SynthID "across multiple modalities of Google models" without naming text, said per-model activation dates are not published, and confirmed there is no public or self-service way to check text from a hosted endpoint.

The citable documentation the developer asked for is still a forum post.

What changes before 2 December?

The grace period ends. The European Commission's Article 50 FAQ, last updated 24 July 2026, says "Article 50 of the AI Act applies as from 2 August 2026" and that "a limited grace period is envisaged only for AI systems placed on the market before 2 August 2026", which comply "only as from 2 December 2026". A provider is whoever puts the system on the EU market under their own name, so buying the model does not move the obligation to the model vendor. We set out the four duties and the two roles in Does the EU AI Act make you watermark your chatbot's text? on 9 August 2026, when the honest answer was that no working text marking existed to build on. Three months later it exists, with a condition attached to it.

Four things to check in your own code. Find every Claude call and remove temperature, top_p and top_k, which is what Anthropic's deprecation note recommends in their place: "Omit and use prompting to guide model behavior." Read the request body your framework builds rather than the arguments you hand it, because two of the eight clients above fill in a field you left alone. Pin the model name. Then log which model wrote which text, because the mark is a property of the model and the output will not tell you afterwards which one produced it.

The migration note states the trade-off outright. "If a request relies on a non-default value, either stop sending the parameter, or keep that request on 2023-06-01." The documented way to keep a temperature of 0.2 for consistent output is to hold your integration on the 2023 API version, which is the one configuration where the mark can go missing without a single error in your logs.

Who does this arrangement suit?

It suits whoever already runs a provenance pipeline. A platform that signs its media assets today adds one more path to a process that exists, with somebody whose job it is. It suits the vendor that wrote the condition down, because a documented limit is a defence, and the one that left the answer in a forum thread is carrying the same risk with none of the paper.

It costs the firm of five whose customer replies come out of a low-code node with 0.9 sitting in the temperature box. Five people, and no compliance department. It costs anyone who lowered temperature in 2025 for consistent output and has not opened a migration guide since. Both are providers under Article 50, and an audit is how they find out.

We built askHermis, which answers 85% of guest queries instantly in 38 or more languages across eight channels. Every one of those answers is generated text that a person reads. The marking duty for a system like that sits with whoever placed it on the market, while the watermark itself is the model vendor's, and whether it is there turns on one field in a request body.

One limit, and it is ours. We read request bodies in source code rather than on the wire, and no vendor has given us detector access, so we cannot tell you whether a single reply any system of ours has sent carries a mark. Nor can any page linked above. If somebody with detector access has a captured request and a verdict on the same text, we would like the pair. If you are scoping a generative feature and want the provenance boundary mapped before it ships, that is the kind of thing we scope.

Which of your own Claude calls still sends a temperature, and which model is on the other end of it?

Common questions

Does setting temperature remove Claude's text watermark?

On six models it can. Anthropic's text watermark page says that if you need the output to carry the watermark you should omit temperature, top_p and top_k, and that a request setting any of them to a non-default value might produce output without the mark. The six models carrying that footnote are Opus 4.6, Opus 4.5, Opus 3, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. On Claude 4.7 and later the parameters return a 400 error instead, so the failure cannot happen quietly.

Which Claude models carry Anthropic's text watermark?

Nineteen, as listed on Anthropic's coverage page on 20 November 2026, across the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag, plus AWS, Google Cloud and Microsoft Foundry. The rollout is recent: the API release notes record the watermark reaching Mythos 5, Opus 4.7 and Sonnet 4.6 on 3 November 2026, and Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 on 16 November 2026.

Can I check whether my AI output carries a watermark?

Not unless a vendor approves you. Anthropic says watermark detection is in private preview for regulators, law enforcement, media, researchers and select enterprises checking their own outputs. OpenAI limits detector access to approved researchers and expert organizations, and Google publishes no public tool for text, only an image, video and audio portal plus an open-source detector for self-hosted models.

Does the Gemini API watermark its text output?

Google does not document it either way. A developer asked on the official Gemini API forum on 5 August 2026 and was told the same day that generated text from the API is not SynthID-watermarked. Two weeks later the same account retracted that after checking with the team and said API text is watermarked. Three later questions about specific models and detection routes went unanswered, and a reply on 16 November 2026 confirmed there is no self-service way to verify a text watermark.

Who has to mark AI-generated text under the EU AI Act, and from when?

Providers of generative AI systems, meaning whoever places the system on the market under their own name. The European Commission's Article 50 FAQ says Article 50 applies from 2 August 2026, and that a limited grace period for the marking and detection duty covers only systems placed on the market before that date, which comply from 2 December 2026. Buying the model from a vendor does not move the duty to the vendor.