The AI startup Anthropic mistakenly billed a free user of Claude $16.6 million. A developer from South Korea, using the pseudonym remy_notes, reported this incident on the social media platform Threads.

He received the first invoice for $1.6 million on July 7. Initially, the programmer thought it was a phishing attempt, but less than a day later, he received a second email demanding payment—this time for $16.6 million.

The sender and payment link pointed to Anthropic's official domain and its payment processor, Stripe. However, remy_notes' dashboard showed zero API usage, and there were no billable keys associated with his account.

On the same day, July 8, the developer's bank card was declined twice for the payment attempts, both of which exceeded the transaction limit. The charges were listed under an American vendor named ANTHROP.

The user checked all processes running on his Mac mini, including eight AI agents, but found nothing that could explain the incident. As a precaution, he canceled his Claude Max subscription and blocked the linked card.

It took four days and around 18 emails to customer support to clarify the situation, the developer wrote. Some responses were automated, coming from an AI agent named Fin.

Eventually, the issue was escalated to the company's privacy department. Meanwhile, the system continued to send repeated requests for the same amount. Ultimately, the bank blocked the developer's card.

Anthropic's Response

On July 12, Anthropic confirmed the error in an email to remy_notes, stating that no money had been withdrawn. The company attributed the issue to a misconfigured auto-replenishment feature and disabled it as a precaution.

Email from Anthropic regarding the error. Source: Threads.

Anthropic linked the problem to the card being blocked, suggesting that the bank reacted to the failed payment attempts. The project team also emphasized that no one had hacked into the programmer's account. However, they could not explain why the system generated a bill for exactly $16.6 million.

Not an Isolated Incident

At the end of June, The Information reported, citing Vaudit, that corporate clients of Anthropic and OpenAI had experienced overcharges. The auditing firm reviewed invoices totaling $34 million across 60 organizations from March to June.

The majority of the invoices were related to the use of Anthropic's Claude Code. Auditors identified excessive charges amounting to about $1.7 million—roughly 5% of the total reviewed.

Vaudit highlighted several systemic errors. Clients were sometimes billed at rates for new models, even though they were actually using a cheaper older version. There were also instances of charges for requests that resulted in errors or yielded no results.

A separate category included "repeat storms"—situations where an AI agent repeatedly attempts a failed task in the background, with each attempt reflected in the bill.

Following inquiries from Vaudit and clients, refunds began to be issued. Approximately 80% of the overcharges were reimbursed by Amazon, Google, Microsoft, Anthropic, and OpenAI.

It’s worth noting that earlier this spring, Anthropic faced user dissatisfaction over limits on Claude Code, prompting the company to double those limits after a deal with SpaceX.