Stunned! A Korean developer received a sky-high Claude bill worth over 500 million yuan, writing code until he went bankrupt?

Korean developer remy_notes’ Anthropic account dashboard shows API usage of $0, yet on July 7–8 he continuously received two Stripe payment failure notifications, with a total amount of more than 500 million New Taiwan dollars, and both pointed to the official billing system.
(Background: Anthropic’s new “Review Dashboard” — review your interactions with Claude, and teach you how to collaborate with AI more intelligently.)
(Additional context: AI fully automatic “hacker extortion attack” exposed! Codenamed JadePuffer, targeting crypto wallets)

Table of contents

Toggle

• From suspected phishing emails to official confirmation
• How big is the billing black box for AI agents?
• In the AI agent era, cost control can’t rely on only one line of defense

The API fee number remy_notes sees in the Anthropic backend is 0, but the payment failure notifications sent by Stripe show a different number: 16,627,739 USD (about NT$534 million).

And this wasn’t even the first bill he received. On the evening of July 7, Stripe already sent the first payment failure notice, for $1,669,875.90 (about NT$53.66 million). The next evening on July 8, the second amount surged to 16.62 million USD, 10 times the day before. Ultimately both transactions showed “payment failed” and no funds were actually deducted from his account, but it still left him unable to sleep at night.

From suspected phishing emails to official confirmation

Afterward, remy_notes published the full account of what happened on Threads: he said he initially thought it was a phishing email, until he checked the sender domain and the payment links and found that both pointed to Anthropic’s official domain. He even called his bank, which told him that these were indeed merchant payment requests issued by Anthropic via Stripe’s billing system, not forged invoices.

According to a report by ZDNet Korea, remy_notes usually runs multiple AI automation systems on a single Mac mini, including 8 AI agents, an automated trading agent, work agents, and other automation scripts. He checked the bots, agents, proxy servers, Codex bots, and Cron jobs one by one, but couldn’t find any API Key that could produce usage at this level.

How big is the billing black box for AI agents?

After the story went to Reddit, discussion split into two camps. One suspects it’s related to the AI agents getting stuck in an endless loop—agents continually calling their own API, causing usage to spiral out of control. The other believes the problem lies with the billing system itself, because even Anthropic would be hard-pressed to imagine that an account could actually spend tens of millions of dollars.

Both theories point to the same thing: no one can explain with 100% certainty how these numbers appeared. Even Anthropic’s own customer support system couldn’t give a clear answer on the spot.

This isn’t a single case either. In June, the auditing startup Vaudit found that in AI service invoices totaling about $34 million, about $1.7 million was charged in excess, and most of it was related to Claude Code. Before that, there had already been reports on GitHub over months about contradictions before and after Anthropic billing.

In the AI agent era, cost control can’t rely on only one line of defense

As of the time of publication, remy_notes has already disabled the credit card used in registration, canceled his Claude Max subscription, and contacted Anthropic customer support to void the invoices and investigate whether his account had been accessed abnormally. As of now, Anthropic has not formally responded, nor explained whether this is a cybersecurity incident, AI hallucination, or simply a technical billing error.

But in an era where AI agents make decisions on their own, this incident highlights the danger of having no safeguards in place (such as setting a per-spend limit, restricting the maximum number of loops per run, auto-stopping on errors, etc.). It’s effectively handing over your credit card number to a piece of code that won’t hit the brakes.