📅 November 3, 2026 • support.claude.com
What happened
Starting November 3, Anthropic applies its invisible text watermark to the outputs of four more Claude models: Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Until now the marking covered only the newest models in the lineup, so this step extends provenance marking across a much larger share of production traffic. Per Anthropic, the rollout is worldwide, there is no opt-out, and the change adds no extra characters or tokens to responses and does not affect pricing. Customers were notified by email on October 28-30. The same marking will reach Claude models served through Amazon Bedrock a few additional days later.
How it works
The watermark is embedded in generated text and remains invisible to readers. Ordinary users cannot detect or remove it, and per Anthropic it does not change the substance or quality of answers. Detection is deliberately restricted: it is available only through a private-preview API for regulators, law enforcement agencies, media outlets, fact-checkers, independent researchers, educational organizations, EU civil society groups and compliance-obligated enterprises. Anthropic positions the mechanism as provenance infrastructure for authorized verification, not as a user-facing feature, which explains why ordinary API customers get no detection access of their own.
The compliance backdrop
The move implements Article 50 of the EU AI Act, which requires providers of AI systems to mark synthetic content in a machine-readable way. That obligation has applied since August 2, 2026, with a transition deadline of December 2, 2026 for systems already on the market. Article 50 is one of the transparency pillars of the Act, alongside the disclosure duties for synthetic media and deepfakes. On the industry side, about 190 signatories have joined the EU transparency Code of Practice, a voluntary framework that many providers treat as the compliance baseline ahead of enforcement.
IMPACT
For enterprises the practical effect is simple: watermarking is becoming a de facto global default. Anthropic frames the change as compliance support for EU customers, but because there is no regional opt-out, outputs from these Claude models are marked everywhere, including the US and Asia. Teams in regulated industries — finance, healthcare, publishing, legal — should assume that model provenance can be verified by authorized third parties and plan their content pipelines accordingly. Audit trails for AI-generated text are quietly turning into standard infrastructure rather than an optional extra, and vendors that mark by default set the reference point competitors will be measured against.
Sources: support.claude.com, artificialintelligenceact.eu, digital-strategy.ec.europa.eu
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
