What happened
Anthropic told API customers by email that it will add text watermarking to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 from November 3, 2026. The watermark is already present on newer models released after August 2, 2026, including Opus 5.5. The rollout covers Anthropic’s API and first-party surfaces; availability on Amazon Bedrock may follow a few days later. Anthropic cites the EU AI Act’s December 2 transition deadline as the reason for retrofitting older models.
Anthropic describes the mark as an imperceptible pattern in word choice. It says the pattern does not change meaning, quality, or readability, adds no characters or tokens, does not alter pricing or request and response formats, and barely slows output. Anthropic also says the watermark carries no information about a user, organization, or conversation. The Decoder reports that the watermark is based on a variant of Google DeepMind’s SynthID text method, which adjusts the randomness used in word selection to create a statistically detectable pattern. Since the mark is embedded at the model level, it applies across hosting platforms including AWS, Google Cloud, and Microsoft Foundry, and customers cannot disable it directly.
There is a reported exception on the two older models Opus 4.6 and Sonnet 4.6. According to The Decoder, Anthropic’s email says the watermark applies only if temperature, top_p, and top_k remain at their default values. Changing any one of those sampling parameters is enough to avoid the mark. Anthropic’s warning is that teams relying on watermarking for compliance should keep these settings at their defaults. The email did not mention the same caveat for Mythos 5 or Opus 4.7; that omission suggests, but does not confirm, that the mark applies regardless of sampling settings on those models.
Anthropic’s watermark-detection API is in private preview, according to the report. The discovery-event summary says preview access is limited to regulators and other eligible organizations. The supplied material gives no price for watermarking or detection, and says the detection API is not generally available. It does not provide independent test results for detection accuracy or a quantified measurement of the watermark’s effect on output.
What the archive adds
This is a new event in a known product area: [Anthropic’s earlier Claude watermark coverage] is reflected in an August 24 Telegram post, which described watermarking in the context of the EU AI Act. New Runtime also covered xAI’s planned default watermarking for Grok API text output, a related but separate rollout: Grok API watermark plan. The new fact here is a specific Claude rollout and a reported sampling-parameter caveat.
Trend signal
This supports the shift toward watermarking as a product and compliance mechanism, but no supplied SITE_PATTERNS thesis specifically covers content provenance or detection. The closest archive precedent is the Grok API watermark plan, which is a related deployment, not evidence for one of the supplied patterns.
Why now
Anthropic attributes the retrofit to the EU AI Act’s December 2, 2026 transition deadline. The company says models released after August 2 already ship with the watermark, while older models are being updated in stages. The announcement therefore concerns both a compliance timetable and an attempt to apply the mechanism to models already in use.
Connections to other events
A recent New Runtime note says xAI planned default invisible watermarking for Grok API text outputs from November 7, 2026, with team-level opt-outs. That provides a parallel example of a provider adding text provenance controls, though the implementation and customer controls differ. The supplied material also reports that OpenAI’s textGrain system activates by default only for ChatGPT and Codex users in the EU, while API users worldwide can switch it on or off; Google DeepMind’s SynthID text has been open-sourced and built into Gemini models.
Why it matters
Teams may need to distinguish between a provider’s watermarking policy and the behavior of a particular production configuration. The reported exception means a stack using Opus 4.6 or Sonnet 4.6 cannot assume that text will carry the mark if it changes a sampling parameter. That matters if an organization communicates that generated text is marked, uses detection as one input to a compliance process, or expects consistent results across models and host platforms.
The rollout also puts limits on what a watermark can establish. Anthropic says it contains no user or organization information, and the detection API remains in private preview. A mark may support detection of generated text, but the supplied facts do not show that it identifies who generated the text, proves how it was used, or replaces broader recordkeeping and governance. Teams should verify model and parameter behavior rather than treat the presence of a feature announcement as evidence that every output is marked.
Limits and uncertainty
The primary public account is The Decoder’s report of an email to API customers; Anthropic’s email is not included in the supplied material. The report’s interpretation that Mythos 5 and Opus 4.7 are marked regardless of sampling settings rests on the absence of a caveat, not an explicit confirmation. The material supplies no independent benchmark of watermark detection, false positives, durability after editing, or effects on text quality. The detection API is not generally available, and rollout timing on Bedrock may differ.
Reader takeaway
Anthropic is extending text watermarking to older Claude models, but the reported behavior on Opus 4.6 and Sonnet 4.6 depends on default sampling settings. If your team relies on the mark, check the exact model and parameters in production, and avoid claiming universal coverage until you have verified it.
Open question
Will Anthropic clarify whether watermarking on Mythos 5 and Opus 4.7 persists when sampling parameters are changed?