Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- The November 3 release notes explicitly add Mythos 5, Opus 4.7 and Sonnet 4.6. The official support table defines product and cloud-platform coverage.
- Amazon Bedrock rollout for these three models is marked for completion by November 10. That deadline does not establish that every account is already covered.
- Sonnet 4.6 output may not reliably carry a watermark with non-default temperature, top_p or top_k. If relying on the mark, Anthropic advises omitting all three, including values inserted by frameworks.
- Watermarking requires no response-handling changes and adds no tokens or hidden characters. Text-detector access is separate from public file-provenance checking.
Which older models gained support? Start with these three
Anthropic's November 3, 2026 Claude Platform release notes confirm that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark on the listed platforms. For content teams and API developers using those models, the immediate change is broader provenance marking. It is not a blanket announcement covering every older model.
A text watermark is a machine-detectable signal embedded in the wording, rather than a visible icon. The official help table separates first-party products, cloud partners and file credentials. Read the relevant column and footnotes: support for file credentials does not establish support for text watermarking.
Bedrock has a rollout deadline, and regional coverage has conditions
Anthropic's help article covers supported-model output in products including Claude, Claude Code, Claude Cowork, Claude Tag and the Claude Platform API, and lists AWS, Google Cloud and Microsoft Foundry as cloud partners. A platform or feature may not support a particular kind of marking, so a product name alone does not settle coverage for every request.
The cloud-partner text-watermark entries for Mythos 5, Opus 4.7 and Sonnet 4.6 carry a footnote: rollout on Amazon Bedrock will complete by November 10. As of this November 4 review, that includes a remaining rollout period. The note gives neither an account-by-account completion time nor an exact time zone.
Anthropic describes regional scope as worldwide, wherever Claude is offered, subject to model, platform and feature support. These sources do not provide a complete region-by-region or account-by-account availability list. Check the conditions of the actual product you use.
Check sampling settings inserted by your Sonnet 4.6 framework
Sampling settings influence how the model chooses its next text. Both the release notes and API guide warn that Sonnet 4.6 requests with a non-default temperature, top_p or top_k may produce text without a watermark, or without a reliable watermark. A model's inclusion in the support table does not guarantee the same result for every parameter combination.
If you rely on the watermark, Anthropic advises omitting all three parameters, including any values set by a framework or client library. Inspect the final request configuration rather than only checking for handwritten settings in application code. Do not invent a universal numeric default, and assess how a change could affect your application's existing sampling behavior.
For example, a hypothetical writing application does not explicitly set temperature, but its wrapper library inserts a value. Inspecting the wrapper's final request is necessary to determine whether it follows the guidance. This illustrates where to check; it is not an API test performed for this article or a guarantee that a single output will be detectable.
Generating a watermark does not grant text-detector access
Anthropic's API guide says the watermark adds no tokens or hidden characters and carries no information about the user or organization. Developers do not need to change requests or response handling for marking. Anthropic says it does not change meaning, quality or readability. That general description does not remove Sonnet 4.6's separate sampling caveat.
Text-watermark detection is currently in private preview. The listed eligible organizations include regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. It also includes enterprises that need to check their own Claude outputs to meet relevant obligations. Registering interest or having a Claude account does not establish detector access. Broader detection API access remains a plan.
The free public Claude Content Checker checks files for Claude-issued Content Credentials. These credentials are signed provenance metadata following the C2PA standard for supported files. This file-checking tool does not make arbitrary text-watermark detection public.
What does a detection result establish?
Finding a supported Claude mark provides a clue that content may have been generated or processed by Claude. People also use Claude to proofread, translate, summarize or convert existing material. A mark alone therefore establishes neither full provenance nor the original author.
A negative result does not prove human authorship. Content created before a model supported marking, short passages, extensive rewriting, translation or mixing with other text can affect detection. Platform, feature and sampling conditions also matter. For files, re-saving, screenshots or format conversion may strip provenance metadata.
Review the model, platform, generation time and edit history together before interpreting a result. Provenance signals replace neither factual verification nor applicable disclosure requirements. Assess the obligations of your particular product and use independently.
What should you change in your content workflow now?
Record the actual model and platform, then check the text column and footnotes in the support table. If using Sonnet 4.6 and relying on its watermark, confirm that all three sampling parameters are omitted from the final request. For Bedrock, follow the November 10 rollout deadline and subsequent clarifications. Organizations needing detection should establish their eligibility and actual approved access first.
For OpenAI output, see the separate OpenAI text watermarking and detector eligibility guide. Providers have different settings and access scopes; one provider's activation path or detector permission does not transfer to another. This article uses Anthropic sources read on November 4. It did not change user settings, request detector access or verify private output.
Sources
Frequently Asked Questions
Will older articles that were already generated gain watermarks automatically?
Expanded model support does not establish retroactive marking of historical text. Anthropic describes embedding a mark during generation and lists content generated before a model supported marking as potentially lacking a detectable mark. Retain the original generation time rather than judging an old article solely by today's model list.
Does this announcement specify text-detection prices or quotas?
The release notes, API guide and help article reviewed here provide no text-detection price or quota that applies to every organization. Check private-preview eligibility, actual access and usage conditions separately. No added watermark tokens does not mean free or unlimited detection.