# What changed on the Claude platform this week, from text watermarks to message threads?

Date: 2026-11-06 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/claude-text-watermark-security-center-message-threads/

**Answer:** Between November 2 and 5, 2026, Anthropic extended its invisible text watermark to more Claude models, added a security center and other admin controls to Claude Enterprise, and launched message threads, a beta that lets long agent conversations send only new messages to the API. The watermark helps show that Claude was likely involved in a text, but it cannot prove who wrote it, and a missing watermark proves nothing.

Anthropic shipped three changes in four days this week, each aimed at a different reader. On November 2, developers got **message threads**, a faster way to run long conversations through the API. On November 3, the company's invisible **text watermark** reached three models released before August: Mythos 5, Opus 4.7 and Sonnet 4.6. On November 5, security teams at large companies got a **security center** and a set of admin controls for Claude Enterprise.

None of these changes what Claude can do for an ordinary user. They change how Claude's output can be traced, how companies police its use, and how quickly agents respond. The watermark deserves the most attention, because it is easy to overestimate.

## Text watermarking: which models, and the exceptions

A **text watermark** is a hidden statistical pattern in the words a model picks. Anthropic announced its method in August and says it adopted it to meet the European Union's AI Act, which since August 2, 2026 asks AI providers to mark generated content in a machine-readable way. Anthropic is one of about 190 signatories of the EU's transparency Code of Practice for content made by AI, and it applies the watermark worldwide because, it says, it has no reliable way yet to limit it by region.

Models launched after August 2 carry the mark from day one. Older models are being added, and this week's update brought in three of them. The current list from Anthropic's Help Center:

| Model | Text watermark | C2PA credentials on files |
|---|---|---|
| Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5, Haiku 5.5 | Yes, on Anthropic's platforms and cloud partners | Yes |
| Mythos 5, Opus 4.7 | Yes; Amazon Bedrock rollout due by November 10 | Yes |
| Sonnet 4.6 | Yes, with an exception (below); Bedrock by November 10 | Yes |
| Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 | Not yet | Yes |

The exception matters for developers. On Sonnet 4.6, a request that changes **temperature**, **top_p** or **top_k** from their default values may produce text without the watermark. These three settings control how random the model's word choices are. Anthropic's advice to anyone who relies on the mark is to leave all three out of requests, including values a framework sets quietly on your behalf.

**C2PA Content Credentials** are a different tool: a signed label in a file's metadata, such as a PNG or JPEG Claude creates, saying Claude made or processed it. Our explainer on [content credentials and AI watermarks](/en/learn/content-credentials-ai-watermarks/) covers that standard.

## How the watermark works

When a model writes, it picks each next word among several candidates. Often two candidates are equally good, and a random number decides between them. Anthropic's watermark, a version of Google DeepMind's SynthID-Text method, replaces that random number with one computed from a secret key and the few words before. Readers see normal text. Someone holding the key can check whether the sequence of choices fits the key, and estimate how likely it is that Claude produced it.

Think of a board game where players move by rolling a die. Now imagine they quietly read their "rolls" from a secret list of digits instead. Each game still looks random, but anyone with the list can compare it to the moves afterwards and see whether the list was used.

Anthropic says the watermark adds no characters and no tokens, costs nothing extra, does not change quality, and carries no information about you or your organization.

## What a watermark can and cannot tell you

| It can | It cannot |
|---|---|
| Signal that Claude was likely involved in a longer passage | Prove that Claude wrote it, rather than edited, translated or summarized someone's text |
| Survive copy and paste, and often light edits | Survive a full rewrite or a paraphrase by another tool |
| Work across Claude products, since it is applied at the model level | Say anything about short texts, which give it too few choices |
| Mark translations done by Claude | Mark code or factual passages much, where only one answer is correct |
| Be checked by those Anthropic gives access to | Detect other companies' AI, which use their own keys and methods |

The most important rule runs in one direction: **a missing watermark proves nothing**. The text may come from an older model, a non-default setting, another AI, a short answer or heavy editing. And a detected mark does not settle authorship either, since a person may have written the ideas and asked Claude to polish them. This differs from commercial AI detectors, which guess from stylistic habits because they do not have the key; our article on [AI detectors in schools](/en/learn/ai-detectors-tell-if-student-used-chatgpt/) explains why their guesses go wrong.

## Claude Enterprise: a security center and new admin controls

The November 5 changes are for companies that run Claude Enterprise. Several are in beta.

| Feature | What it does | Status |
|---|---|---|
| Security center | One page comparing your Claude settings for chat, Cowork and Claude Code with Anthropic's recommended configuration, with a downloadable report for auditors | Beta |
| Inference hooks | Send prompts, and tool calls before they run, to the company's data-loss-prevention tools for a check | Generally available |
| Centralized management | One admin console at admin.claude.com for sign-on, user directories and domains across several organizations | Beta; setup cannot be undone |
| Sign a member out everywhere | One Admin API call ends a member's sessions on every device | Beta |
| Self-serve IP allowlist | Admins limit access to approved network addresses themselves | Beta |
| Block outside accounts in Claude Desktop | Stops personal Claude accounts on managed computers, to curb unapproved "shadow IT" | Beta |

**Data loss prevention (DLP)** tools scan outgoing data for things like customer numbers or source code. Anthropic is explicit that the security center is guidance based on its own recommendations, not a certification of compliance with any law or standard.

## Message threads: smaller requests for long agent loops

The Messages API is **stateless**: each request normally resends the whole conversation, including every earlier message and tool result. For a coding agent that has run for hours, that can mean hundreds of thousands of tokens uploaded again on every turn.

With message threads, the first request sends the full conversation and opens a thread. After that, every request carries just the turns added since, together with the id of the previous response. The server holds the thread in encrypted memory, not on disk, for 5 minutes after the last request and deletes it within an hour. Anthropic reports that with threads switched on for most Claude Code users, the median wait before the first token on requests of 500,000 to 1 million tokens fell by about 35 percent.

| Request | Without a thread | With a thread |
|---|---|---|
| Turn 1 | Full conversation | Full conversation, plus `"thread": {"type": "create"}` |
| Turn 50 | All 50 turns again | Only turn 50, plus `"thread": {"type": "continue", "previous_message_id": "..."}` |
| Tokens billed | Full conversation | Still the full conversation: the model reads all of it |

Three cautions. Threads save time, not money, because Claude still reads every token. They are best-effort, so your application must keep its own copy and resend everything when a thread has expired. And the documentation says threads are not eligible for zero data retention or HIPAA, which rules them out for some health and regulated workloads. Prompt caching still has to be requested separately, as described in our [prompt caching guide](/en/learn/prompt-caching-cut-api-costs/).

> **Try it:** If you call the Claude API through a framework, log one outgoing request and check whether it sets temperature, top_p or top_k; on Sonnet 4.6 those settings can switch the watermark off.

## Can you check whether a text was written by Claude?

Not yourself, for now. Anthropic's watermark detector is in private preview, open to groups the EU rules name, such as regulators, police, media, fact-checkers, researchers, schools and civil society organizations, plus companies that need to check their own Claude output to meet the AI Act. Others can register interest through a form Anthropic links from its Help Center, and the company says access will widen over time. For files, a free Claude Content Checker reads the C2PA credentials.

Even with access, a result is a probability that Claude was involved, not a verdict. Treat it as one clue alongside others, such as drafts, version history or a conversation with the author, and never as proof on its own that someone did or did not use AI.

## Sources

- Claude Platform release notes: https://platform.claude.com/docs/en/release-notes/overview
- How Claude marks AI-generated content: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
- How Claude's text watermark works: https://www.anthropic.com/news/claude-text-watermark
- More visibility and control over Claude Enterprise for security teams: https://claude.com/resources/articles/claude-enterprise-security-updates
- Make long-running agents faster with message threads: https://claude.com/resources/articles/message-threads
- Message threads: https://platform.claude.com/docs/en/build-with-claude/message-threads

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/claude-text-watermark-security-center-message-threads/, no commercial use without permission, and keep the same licence.