Anthropic has reportedly told API customers that Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 will begin embedding text watermarks on November 3, 2026. The change was outlined in an October 28 customer email, according to The Decoder’s October 31 report.
The rollout reportedly covers Anthropic’s API and cloud partners, although Amazon Bedrock may take a few extra days. Two models also have a consequential exception: Opus 4.6 and Sonnet 4.6 will reportedly watermark text only when three sampling parameters remain at their defaults.
Anthropic’s public support table still lists all four models as lacking text watermarks while supporting C2PA credentials for files. The November 3 schedule and sampling exception come from the reported customer email, not a dated public Anthropic announcement. That separates the planned rollout from the capabilities currently documented in the official support page.
The Four Models and Their Documented Status
Anthropic’s support documentation distinguishes text watermarks on its own products, text watermarks on cloud partners, and C2PA Content Credentials attached to supported files.
For the four models named in the email, the available support table and reported plan line up as follows:
| Model | Current documented text watermark status | C2PA file credentials | Reported November 3 change |
|---|---|---|---|
| Claude Mythos 5 | Not yet supported | Supported | Text watermarking; no sampling exception reported |
| Claude Opus 4.7 | Not yet supported | Supported | Text watermarking; no sampling exception reported |
| Claude Opus 4.6 | Not yet supported | Supported | Text watermarking at default sampling settings |
| Claude Sonnet 4.6 | Not yet supported | Supported | Text watermarking at default sampling settings |
Sources: Anthropic’s support table and The Decoder’s account of the email.
This looks like a documentation timing gap, not evidence that the two sources describe incompatible policies. Anthropic’s support page explicitly says it is adding marking support to models released before August 2, 2026.
Still, developers should keep the distinction intact. A scheduled change is useful for planning, while the support table describes published availability. Neither alone confirms the exact moment a particular cloud deployment switches over.
The Sampling Exception Creates a Configuration Trap
Opus 4.6 and Sonnet 4.6 reportedly apply the watermark only when temperature, top_p, and top_k all remain at their default values. According to The Decoder, changing any one of these settings is enough to avoid the watermark.
These parameters control how the model selects its next output:
temperatureadjusts randomness in selection.top_prestricts the candidate pool according to cumulative probability.top_klimits the number of candidate tokens considered.
They are ordinary generation controls that developers may tune for writing, structured output, or other application requirements. That makes the exception operationally important even for teams with no interest in bypassing provenance marking.
The email’s warning, as quoted by The Decoder, is explicit:
Source: The Decoder“If you rely on the watermark for your own compliance, keep these settings at their defaults.”
The report mentions no equivalent exception for Mythos 5 or Opus 4.7. That is a difference in the reported guidance, not proof that every possible configuration on those models has been independently checked.
The practical implication is that model identity alone will be insufficient to establish expected watermark coverage on Opus 4.6 and Sonnet 4.6. The effective sampling configuration matters too.
I’d treat this as a configuration-management issue. A team could select a supposedly watermarked model, then lose that coverage because a framework or application layer supplies a non-default sampling value. Developers should inspect the settings actually sent through each integration instead of relying on a model-name allowlist.
How Claude’s Text Watermark Works
Anthropic described its approach in “How Claude’s text watermark works,” published August 14, 2026. It uses a version of Google DeepMind’s SynthID-Text method.
The mechanism operates during generation. A language model often has several plausible choices for what comes next. Anthropic says its watermark changes the source of randomness used to select among those choices, using a key and preceding text to produce a statistically recognizable pattern.
The result is embedded in the word sequence itself. It is not a visible label, hidden character, or extra block of metadata appended to the response.
According to Anthropic:
- Watermarking adds no extra tokens and does not increase token costs.
- Readers should not be able to distinguish watermarked text by appearance.
- The watermark contains no identifying information about a person, organization, or conversation.
- Its internal testing found no effect on content, creativity, or readability.
Those quality findings are Anthropic’s assessment. They are not an independent measurement of the four-model rollout.
One useful distinction follows from the design: the watermark signals likely model involvement, not the identity of the customer who requested the text. It is not described as a customer-specific tracking code.
Because the pattern resides in the text, copying and pasting can carry it into another application. That differs from a file credential, which depends on provenance metadata attached to a supported file.
Cloud Coverage Does Not Mean Identical File Support
Anthropic’s support documentation says supported models carry text watermarks across its own products and cloud partners worldwide.
Its first-party list includes Claude Platform, the Claude apps, Claude Code, Claude Cowork, and Claude Tag. Cloud coverage includes AWS, Google Cloud, and Microsoft Foundry. The watermark is applied at the model level, so supported text marking is intended to follow the model across these surfaces.
The reported November 3 expansion follows that broad approach, with potentially different deployment timing. The Decoder specifically says Amazon Bedrock may take a few extra days.
For teams using multiple providers, that creates a temporary possibility: the same named model could have different watermark availability depending on where and when a request is served. A deployment record should therefore include the provider and rollout status, not only the model version.
C2PA Applies When Claude Creates Supported Files
C2PA Content Credentials are a separate mechanism. When Claude generates a supported file, such as a PNG or JPEG, it can attach signed provenance metadata indicating that Claude processed it.
Anthropic says this applies where the platform offers Claude’s file-generation features. Its documentation names the Claude apps and Claude Platform API, including Claude Platform on AWS and Claude in Microsoft Foundry.
That wording matters. Text watermark support through a cloud provider does not automatically mean every service on that provider supports Claude-issued file credentials. In particular, the broad text-watermark coverage statement should not be read as a promise of identical C2PA functionality on Google Cloud.
The four models’ existing C2PA support therefore does not remove the significance of the text rollout. They cover different outputs and use different verification methods.
Detection Is Restricted and Statistical
Anthropic’s text-watermark detection API remains in private preview. Its support page lists eligible organizations including regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups. Enterprises with relevant compliance obligations can also qualify.
Anthropic says it plans to expand access over time. This is not currently a broadly available public text-checking service.
Frequently Asked Questions
3 questions
1Will the Watermark Increase Token Costs?
Anthropic says its text watermark adds no tokens and does not make generation more expensive.
