Anthropic will add watermarks to four more Claude models

Which models will get watermarks

Anthropic initially planned to add watermarks to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. In an update on November 4, the company confirmed that Opus 4.6 would get a watermark on November 16, although it had originally been scheduled for an update on November 3. Anthropic did not give a reason for the delay.

The new email does not mention Mythos 5, Opus 4.7 or Sonnet 4.6. This suggests their updates went live on November 3 as planned. Opus 4.5, Sonnet 4.5 and Haiku 4.5 were not included in the earlier announcement.

All four models in the new rollout share the same condition: watermarking applies only when the temperature, top_p and top_k settings are left at their defaults. In the first rollout, that restriction applied to only two of the four models—Opus 4.6 and Sonnet 4.6.

Newer models, such as Claude Opus 5.5, already have watermarks. Deployment on Amazon Bedrock may take a few extra days.

How the watermark works

Anthropic describes the watermark as a pattern in word choice that readers cannot notice. The company says it does not change the text’s meaning, quality or readability, add characters or tokens, affect the price or format of requests and responses, and only slightly slow generation. The company also says the watermark contains no information about the user, their organization or their conversations with Claude.

On Opus 4.6 and Sonnet 4.6, watermarking is enabled only if developers leave temperature, top_p and top_k at their default values. These settings control how much randomness is used in word choice. Changing any one of them is enough to bypass the watermark. Anthropic warned customers that if they use watermarking to meet their own requirements, they need to keep the default settings.

The email does not say whether the same condition applies to Mythos 5 and Opus 4.7. This may mean watermarking works on those models with any generation settings. At the same time, Anthropic considers Opus 4.6 its best model for writing—even compared with Opus 5.5.

To test the watermark, the company is running an API in a closed preview. More information is available on its content watermarking page.

Why older models are being marked

Anthropic links the updates to older models to a December deadline set by the EU AI Act. All models released after August 2, 2026, received a watermark from the outset; the company is updating earlier models in stages.

Because the watermark is built into the model, it applies on every platform where the model is hosted, including Amazon Web Services, Google Cloud and Microsoft Foundry. Customers cannot turn it off.

Anthropic uses a version of Google DeepMind’s SynthID text method. It adjusts randomness in word choice to create a statistically detectable pattern in the text.

In October 2026, OpenAI launched its own text-marking system, textGrain. It is enabled by default only for ChatGPT and Codex users in the EU. API customers worldwide can turn it on or off themselves; it is off by default. Google DeepMind open-sourced the text version of SynthID and built it into Gemini models.

Debates over accuracy and legal risks

Whether watermarks reduce text quality remains disputed. Blogger John Gruber argues that synonyms are never fully identical in meaning. In his view, choosing words according to a watermark key rather than their fit with the meaning inevitably reduces accuracy.

Artificial Lawyer separately points to potential problems for law firms. Clients who have banned the use of AI may find a watermark in a document, and multiple watermarks layered on top of one another could make it harder to work with.

Dataist's take: Model-level watermarking makes it easier to apply across platforms, but does not guarantee that text can be reliably verified under every setting: on Opus 4.6 and Sonnet 4.6, the watermark depends on generation parameters. The debate over whether watermarks affect accuracy in word choice also remains open.