---
type: "post"
id: "anthropic-to-apply-its-eu-ai-act-text-985b99"
slug: "anthropic-to-apply-its-eu-ai-act-text-985b99"
title: "Anthropic to apply its EU AI Act text watermark to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3"
description: "Anthropic is reported to be extending text watermarking to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3."
retrieval_nugget: "Anthropic is reported to be extending text watermarking to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3. For teams using those models, the change matters because generated text will carry a machine-detectable signal across Anthropic’s services, while the detection tool remains restricted and the signal is not conclusive."
published_at: "2026-11-01"
updated_at: "2026-11-06"
record_date: "2026-11-01"
date_kind: "published_at"
topics: ["anthropic","claude","google"]
entities: ["Anthropic","Claude","Mythos 5","Opus 4.7","Opus 4.6","Sonnet 4.6","SynthID-Text","EU AI Act"]
related_patterns: []
novelty_label: "same_event"
what_changed: "Same event as Anthropic will extend its invisible text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on November 3"
prior_refs: ["memory:5c344136-2b20-82a0-ae11-fe88103cddfe","memory:753e1ba4-5206-8d9f-ab1d-c976da2661ad","memory:e429747f-3b98-80b9-af27-f8192f53b1fb"]
source_url: "https://maccus.fr/en/claude-text-watermark"
source_urls: ["https://maccus.fr/en/claude-text-watermark"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:anthropic-to-apply-its-eu-ai-act-text-985b99"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/anthropic-to-apply-its-eu-ai-act-text-985b99/","markdown":"https://newruntime.com/posts/anthropic-to-apply-its-eu-ai-act-text-985b99.md","json":"https://newruntime.com/posts/anthropic-to-apply-its-eu-ai-act-text-985b99.json"}
---

# Anthropic to apply its EU AI Act text watermark to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3

## Retrieval answer

Anthropic is reported to be extending text watermarking to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3. For teams using those models, the change matters because generated text will carry a machine-detectable signal across Anthropic’s services, while the detection tool remains restricted and the signal is not conclusive.

## Summary
Anthropic is reported to be extending text watermarking to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3. For teams using those models, the change matters because generated text will carry a machine-detectable signal across Anthropic’s services, while the detection tool remains restricted and the signal is not conclusive.

## What happened
MACCUS reports that Anthropic emailed organizations using at least one of the four models through its API, Claude Platform. A screenshot of the email became public on October 28; MACCUS published its account on October 31, 2026. The email reportedly says text watermarking will begin on November 3 for Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, using the same system already applied to newer models. The source says deployment through Amazon Bedrock may take a few additional days. The date was not listed on Anthropic’s help page or platform release notes as of the report.

The change is part of a wider rollout. MACCUS says Anthropic’s help page, updated October 28, lists 16 models in a table: as of October 31, nine already had text watermarking, while the four newly named models were among those listed only with Content Credentials metadata. After the reported November 3 expansion, Opus 4.5, Sonnet 4.5 and Haiku 4.5 would still have no specific watermarking date. The source notes that Sonnet 4.5 is scheduled to leave the Claude API on November 30, while Haiku 4.5 remained available through the API after its planned removal from some consumer plan model pickers.

Anthropic’s described method is based on a version of SynthID-Text. When multiple words are plausible, a secret key and preceding words influence which one is selected. The method adds no visible characters, hidden characters or extra tokens. MACCUS says it applies at the model level across the API, Claude apps, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry. Anthropic says it applies wherever Claude is offered because it cannot yet be limited durably by region. The source reports no user setting to disable it.

The detection API is in private preview. MACCUS says Anthropic lists regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil-society groups and enterprises needing to verify compliance as eligible groups. Access may be requested, and Anthropic plans to expand it over time. The report places the rollout against EU AI Act Article 50 requirements and says the Digital Omnibus gives systems placed on the market before August 2, 2026 until December 2, 2026 to comply with marking requirements.

## What the archive adds
[Anthropic will extend its invisible text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on November 3](https://newruntime.com/posts/anthropic-will-extend-its-invisible-text-watermark-to-d10caa/) already covers the rollout announcement. A specific operational caveat comes from [The Decoder’s report](https://the-decoder.com/anthropics-text-watermark-is-easy-to-bypass-on-several-older-claude-models/) on the same email: on Opus 4.6 and Sonnet 4.6 the watermark applies only when the temperature, top_p and top_k sampling parameters are left at their default values. MACCUS does not mention this exception. The earlier [explanation of Anthropic’s text watermarking for EU AI Act compliance](https://newruntime.com/posts/anthropic-explains-claude-text-watermarking-for-eu-ai-act-compliance/) supplies the context; this is a new fact on a known story.

## Trend signal
This supports [agent-ready software exposes capabilities](https://newruntime.com/patterns/agent-ready-software-exposes-capabilities) only weakly: model outputs are being made machine-detectable, but the report describes provenance marking rather than a composable capability or permission boundary. The stronger signal is that regulatory compliance is becoming a product behavior that spans model interfaces and cloud distribution, not merely a policy statement.

## Why now
The source connects the November rollout to the EU AI Act’s marking requirement. It says Article 50 became applicable on August 2, 2026, while the Digital Omnibus gives systems placed on the market before that date until December 2 to comply. The reported November 3 schedule would put these four models about a month ahead of that deadline.

## Connections to other events
An earlier Telegram post described Anthropic’s watermarking approach in the context of EU AI Act compliance; this report adds the expansion to more Claude models and the reported November date. The archive also records that xAI planned default invisible watermarking for Grok API text outputs from November 7, 2026, with team-level opt-outs. That is a parallel vendor response to the same regulatory context, not evidence that the two systems work identically.

## Why it matters
Teams that rely on stable text characteristics from these Claude models may see a change without modifying their application code: the reported marking happens during model generation and applies across listed interfaces and cloud partners. It is not a conventional metadata field that a team can inspect in a response, and ordinary users are not given the detector. That makes it important to distinguish internal governance from proof about any particular passage.

For compliance teams, the rollout may help establish a machine-readable signal, but it does not settle whether a piece of text was written by Claude. The source says proofreading, translation, summarization or subsequent editing can affect interpretation; short text, code and factual passages may produce weaker signals. Teams should not make high-stakes authorship decisions from the watermark alone, nor assume that unmarked text is human-written. They should verify their exact model and serving path, and track Anthropic’s official documentation and detection-access terms as the reported rollout approaches.

## Limits and uncertainty
The central rollout claim rests on a screenshot of an email reported by MACCUS, rather than a public Anthropic announcement confirming the November 3 date. The source says the help page did not list that date as of October 31, and the Bedrock schedule may lag. Its account of model coverage and legal timing is a secondary report, not legal advice. Anthropic describes detections as not fully conclusive; MACCUS says the source provides no false-positive rate. The material also does not establish the practical impact on output quality or exact detector access conditions, and the sampling-parameter exception for Opus 4.6 and Sonnet 4.6 is reported by The Decoder, not by MACCUS.

## Skill to build now
Create a small model-and-provenance inventory for Claude deployments: record the model identifier, whether traffic goes through Anthropic, Bedrock, Google Cloud or Microsoft Foundry, and where generated text is stored or passed downstream. This week, check the provider’s documentation and avoid treating a watermark result as proof of authorship. That inventory will make it easier to assess the reported rollout and update user-facing disclosures if your workflow depends on generated text.

## Reader takeaway
The reported change extends Claude text marking to four more models, but the mark is a probabilistic signal, not an authorship verdict. Builders should check their deployment path and keep compliance decisions separate from claims about who wrote a particular passage.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)