Anthropic extends invisible text watermarking to Mythos 5 and Opus 4.7 API outputs
Anthropic is extending invisible text watermarking to API outputs of Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3, ahead of an EU AI Act deadline in December.
From November 3, Anthropic is extending invisible text watermarking to API outputs of Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, with implementation landing ahead of an EU AI Act deadline in December. The technique embeds a signal invisible to the human eye in generated text to mark content as model-produced.
The expansion drew objections from paying users over asymmetric watermark treatment, with the main concern that not all calling paths receive equivalent marking. As the EU AI Act approaches its obligations on systemic risk assessment and incident reporting for general-purpose models, traceability in output is becoming part of compliance for frontier labs.
In the same period Anthropic disclosed complete third-quarter figures for the first time in its amended prospectus: revenue of $16.96 billion against $1.38 billion a year earlier, and a net loss of $23.29 billion including $21.9 billion of non-cash charges. As of September 30 the company held $95.6 billion in cash and $36.3 billion in short-term investments, with total contract commitments of about $895.4 billion.