Anthropic acknowledged a billing error on July 12, confirming that its system generated phantom invoices of $1.67 million and then $16.6 million against a South Korean developer on the free tier who had never spent a dollar on the API and had no payment card registered to the account, though no money was collected. The developer, posting as remy_notes, discovered the first invoice—a "payment unsuccessful" notice for $1,669,875.90—on the night of July 7, and screenshots showed his KB Kookmin check card declining two overseas charge attempts on July 8 at 6:01 PM and 7:41 PM from a US merchant listed as "ANTHROP." Anthropic told him the incident "was not the result of unauthorized access" but has not disclosed the root cause. Separately, between March and June of this year, AI billing audit startup Vaudit reviewed $34 million in AI invoices submitted by 60 enterprise customers and found approximately $1.7 million in mistaken overcharges—a billing error rate of roughly five percent—with the majority tied to Anthropic's Claude Code product. Vaudit's client roster included Panasonic, HP, and Honda, and CEO Michael Hahn described recurring billing failure patterns including customers charged at premium rates for older models, billed for failed requests, and "retry storms" generating unauthorized charges.