🚨BREAKING: A Korean developer's Anthropic invoice grew from $1.6 million to $16.6 million overnight. His dashboard said he owed zero dollars the entire time.

Here's what actually happened. The developer, posting on Threads under the handle remy_notes, said his first instinct was that it had to be phishing. Then he checked the sender domain and the payment link. Both were real, verified Anthropic infrastructure. His bank ended up blocking the charge attempts, not because they caught fraud, but because $16.6 million exceeded his card's per-transaction limit.
This isn't an isolated glitch either. A billing-audit startup called Vaudit reviewed $34 million in AI invoices across 60 companies between March and June, including Panasonic, HP, and Honda. It found roughly $1.7 million in overcharges, most of it tied to Claude Code. Common causes: customers billed for a pricier model when they'd actually used a cheaper one, and AI agents stuck in retry loops quietly running the meter without anyone noticing.
Anthropic has said publicly that it hasn't seen evidence supporting the overcharging claims. At the same time, roughly 80% of the disputed charges Vaudit flagged did get refunded, by Anthropic, OpenAI, and the major cloud providers together.
Enterprise AI billing has gotten complicated enough that even the company sending the invoice can't always explain it in real time. That's the actual story here, not one wild number, but how many steps sit between usage and what shows up on a bill.