# Claude Text Watermark Expands: Supported Models, Detection and API Limits

Anthropic has confirmed watermarking for three older models, but the rollout details matter more than a simple yes-or-no detector result.

By Paras Madan | 2026-11-04


Anthropic's November 3 release notes confirm text watermarking for Claude Mythos 5, Opus 4.7 and Sonnet 4.6. The change extends an existing marking system to older models. Developers do not need to change normal API handling, but anyone relying on the watermark needs to check model coverage, sampling settings and platform rollout.

## Which Claude models now carry the text watermark?

The [dated platform release note](https://platform.claude.com/docs/en/release-notes/overview) names three models: Mythos 5, Opus 4.7 and Sonnet 4.6. This is an expansion of coverage, rather than the first introduction of Claude watermarking.

There is an important distinction from earlier reporting about customer notices. **Opus 4.6 is not in the November 3 release note**, and its text-watermark cells remain blank in Anthropic's current support table. Its file Content Credentials column is checked. Those are different capabilities, so Opus 4.6 should not be counted as a confirmed text-watermark rollout on that evidence.

Anthropic's [current coverage table](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content) lists the three newly named models across first-party and cloud-partner output. A footnote says their Amazon Bedrock rollout will complete by November 10. A supported model name therefore does not, by itself, establish that every endpoint has finished rolling out the change.

## How the watermark works

The watermark is a statistical pattern in the model's word choices. It is not a hidden character pasted into the answer or a visible label at the end of a paragraph.

In its [technical explanation](https://www.anthropic.com/news/claude-text-watermark), Anthropic describes a version of SynthID-Text. When several next words would work, a secret key and preceding words help determine the random choice. A detector with the key can look for evidence of that pattern across a passage.

Anthropic says the approach adds no tokens, changes no pricing and contains no information identifying the user, organisation or conversation. It also reports no practical effect on output quality. These are the provider's claims, not results from testing performed for this article.

That mechanism helps explain why a code snippet and a long free-form essay are different detection problems. Exact code and tightly constrained factual answers offer fewer interchangeable choices. A lack of a detectable pattern in those outputs does not establish human authorship.

## The Sonnet 4.6 API setting to check

The [API documentation](https://platform.claude.com/docs/en/build-with-claude/text-watermark) gives one specific caveat: Sonnet 4.6 output might not carry the watermark when temperature, top_p or top_k is set to a non-default value.

If a product depends on marking, Anthropic advises omitting all three parameters, including values inserted by a framework or client library. That last detail is useful. A team can leave its own application code unchanged while a wrapper silently supplies a sampling setting.

For ordinary integrations, Anthropic says no request or response-handling changes are needed. For a workflow that treats watermark coverage as a requirement, the practical check is more specific: record the model, serving platform and effective request parameters together. A model label alone is too coarse.

## Can anyone detect Claude-generated text?

Anthropic's watermark detector remains in private preview for eligible organisations and certain enterprises. It is not a general public text checker. The public Claude Content Checker verifies supported file credentials, which should not be confused with detecting the text watermark.

A detected mark suggests Claude processed the material. It does not establish that Claude originated every idea or wrote the entire document. Conversely, short passages, heavy editing and unsupported outputs can lack a detectable signal.

**Provenance needs more than a detector verdict.** A publishing system should still keep the original output, the edits made afterwards and the review history. The watermark can add evidence to that record; it cannot reconstruct it.

The useful question now is how reliably products preserve that distinction as coverage expands. More marked output will only improve transparency if the people interpreting it understand what the mark actually establishes.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product changes and engineering details into clear stories developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).