---
title: Claude now watermarks the text of all its models: what changes and the temperature catch
url: https://hernandez.systems/en/articles/claude-now-watermarks-the-text-of-all-its-models-what-changes-and-the-temperature-catch
lang: en
type: NewsArticle
section: AI
published: "2026-11-17T14:35:03+00:00"
modified: "2026-11-19T16:11:54+00:00"
publisher: Osky Networks
image: https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-11-17/e76-portada-30373804.png
alternate_es: https://hernandez.systems/es/articulos/claude-ya-marca-el-texto-de-todos-sus-modelos-que-cambia-y-el-detalle-de
---

# Claude now watermarks the text of all its models: what changes and the temperature catch

> Anthropic has finished adding its invisible watermark to the Claude models that didn't have it yet. In its November 3 release notes it added it to Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6, and in the November 16 ones to Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5. With that, every model listed in its support table now watermarks the text it generates, from Opus 3 to Fable 5.2. If you use Claude in a bot, a content workflow or an agent, you'll barely notice the change. But there is one technical detail that does matter: on several 4.x models, changing temperature, top_p or top_k can leave the text without a watermark.

Anthropic has finished adding its invisible watermark to the Claude models that didn't have it yet. In its [November 3](https://platform.claude.com/docs/en/release-notes/overview#november-3-2026) release notes it added it to Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6, and in the [November 16](https://platform.claude.com/docs/en/release-notes/overview#text-watermark-claude-opus-4-6-and-4-5-models) ones to Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5. With that, every model listed in its [support table](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) now watermarks the text it generates, from Opus 3 to Fable 5.2. If you use Claude in a bot, a content workflow or an agent, you'll barely notice the change. But there is one technical detail that does matter: on several 4.x models, changing `temperature`, `top_p` or `top_k` can leave the text without a watermark.

## What Anthropic announced

Both release notes say the same thing for different models: the text they generate "carries Anthropic's text watermark" on the platforms listed in the support article. The November 16 note adds a warning: if a request to those models sets `temperature`, `top_p` or `top_k` to a non-default value, the text might not carry it.

This isn't the first time it has shown up in the changelog. On September 1, when Claude Fable 5.1 and Claude Mythos 5.1 launched, [Anthropic noted](https://platform.claude.com/docs/en/release-notes/overview) that their text already came watermarked and that the image, video and audio files they produce with the code execution tool carry Content Credentials (C2PA) when downloaded through the Files API. November's update fills in the catalog going backwards.

According to the [support article](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content), the watermark applies to the text of supported models everywhere: the API (Claude Platform), the Claude apps, Claude Code, Claude Cowork and Claude Tag, and also when you use Claude through AWS, Google Cloud or Microsoft Foundry. There are two caveats in that table: on Amazon Bedrock, rollout "may take a few more days" for Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5, and Opus 3 only carries it on Anthropic's own platforms, not on third-party clouds.

![Chart showing the Claude models that already watermark their text: those launched after August 2, 2026, those added afterwards, and those that may lose the watermark if temperature, top_p or top_k are changed](https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-11-17/e76-modelos-80299879.png)

## Why now: the European AI Act

Anthropic says it plainly in its [technical explanation](https://www.anthropic.com/news/claude-text-watermark): it's doing this to comply with the European Union's AI Act. Since August 2, 2026, the EU requires anyone offering AI in its market to mark generated content, and Anthropic signed the Code of Practice on transparency of AI-generated content in July, along with "around 190" signatories according to its own text (the European Commission [published its endorsement of the code](https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content)).

The commitment described in the [support article](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) has two parts. Models launched in the EU since August 2 carry the watermark from day one; in the changelog that covers Fable 5.1, Mythos 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.2 and Mythos 5.2. Earlier ones get a transition period, which is exactly what Anthropic was wrapping up in November.

One fact that matters if you operate outside Europe: the watermark is applied worldwide. Anthropic explains it does it this way "because we don't yet have a durable way to limit it by region" and that it will keep evaluating alternatives. If your bot serves customers in Mexico, its text is watermarked too.

## How it works

The technique is a version of SynthID-Text, the method Google DeepMind published in [Nature in 2024](https://www.nature.com/articles/s41586-024-08025-4). When a model writes, at each step it picks the next word from several candidates. Often two options are equally good ("cloudy" or "gray") and a random number decides. The watermark doesn't add anything to the text: it changes where that randomness comes from. Instead of just any random generator, it uses a secret key and the previous words. Whoever holds the key can check whether the sequence of choices matches and calculate how likely it is that Claude was involved.

![Diagram of how Claude's text watermark works: the model chooses between equally good words and, with the watermark, the randomness comes from a secret key; below, the cases where it leaves almost no trace](https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-11-17/e76-funciona-0604df3e.png)

That's where its limits come from, which [Anthropic acknowledges](https://www.anthropic.com/news/claude-text-watermark):

- **Where there are no options, there's no watermark.** An exact fact, the result of a sum or almost any code leave little room. In code it may show up in the comments, not in the logic.
- **Short texts and light corrections** don't give enough signal. If Claude only fixed the spelling of a human-written text, the watermark lives in a handful of changes.
- **A full rewrite erases it.** A light edit "probably" doesn't.
- **Translations do carry it**, because Claude picks every word.

## What changes if you build with the API

For most people, nothing. The [documentation](https://platform.claude.com/docs/en/build-with-claude/text-watermark) says the watermark doesn't change meaning, quality or readability, that it doesn't add tokens or hidden characters, and that it carries no information about you or your organization. Anthropic adds in its [explanation](https://www.anthropic.com/news/claude-text-watermark) that it costs nothing extra and that its effect on speed is negligible. You also don't need to clean weird characters out of your responses or change how you process them.

The fine print is the sampling parameters. The documentation has a section that applies to Claude Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 and Opus 3: if you need your output to carry the watermark, don't send `temperature`, `top_p` or `top_k`. And it warns about something easy to overlook: your framework or client might be sending them on its own, so you have to check the actual request body.

| Situation | What happens to the watermark |
|---|---|
| Model launched since August 2026 | Included from launch |
| 4.x model on the list, sampling untouched | Included |
| 4.x model on the list with `temperature`, `top_p` or `top_k` changed | Might not be included |
| Very short text, code or light correction | Too little signal to detect it |

My take: this hits bot builders in two ways. If your company is bound by the AI Act and counts on Claude's watermark to comply, check that your integration with 4.x models isn't sending a `temperature` of 0.7 inherited from some old example. And if anyone thought raising or lowering the temperature was a way to "remove" the watermark, the documentation doesn't guarantee that: it says the output "might not carry it," not that it gets eliminated.

## What you still can't do

Verify your own text. The [watermark detector](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) is in private preview and is only used by regulators, law enforcement, media, fact-checkers, researchers, educational institutions, EU civil society organizations and some companies that must review their own outputs to comply with the law. Anthropic has a form to request access and says it will gradually open it up. What is public is the [Claude Content Checker](https://claude.com/check-content), but that one checks Content Credentials (C2PA) in files, not the text watermark.

It's also worth understanding what the watermark proves and what it doesn't. According to Anthropic, finding it indicates that Claude "may have" generated or processed the content, but it doesn't confirm who the original author is (if you asked Claude to translate or summarize something, the result comes out watermarked). Not finding it doesn't prove the text is human either: it may have come from a model before it had the watermark, been heavily edited or been too short. And it doesn't change who owns the text or who is legally responsible for it.

According to the [support article](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content), anyone deploying Claude in their own product has to assess on their own what Article 50 of the AI Act requires of them. Anthropic's watermark helps, but it doesn't replace that analysis.

I didn't test the detection or measure the effect of sampling: all of this comes from Anthropic's release notes, documentation and official explanations.

## Sources

- [Claude Platform: November 3, 2026 release notes](https://platform.claude.com/docs/en/release-notes/overview#november-3-2026)
- [Claude Platform: November 16, 2026 release notes](https://platform.claude.com/docs/en/release-notes/overview#text-watermark-claude-opus-4-6-and-4-5-models)
- [Claude Platform: Text watermark](https://platform.claude.com/docs/en/build-with-claude/text-watermark)
- [Claude Help Center: How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content)
- [Anthropic: How Claude's text watermarking works](https://www.anthropic.com/news/claude-text-watermark)
- [European Commission: strong backing for the Code of Practice on transparency of AI-generated content](https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content)
- [Nature: Scalable watermarking for identifying large language model outputs (SynthID-Text)](https://www.nature.com/articles/s41586-024-08025-4)
