A South Korean software developer says Anthropic’s billing system attempted to charge him $16.6 million for the Claude API despite being on the free plan with no usage, in an incident that spread widely online this week.
According to the user, who posted screenshots on Threads and spoke to ZDNet Korea, he received a “payment unsuccessful” invoice for about $1.67 million on July 7, followed a day later by a second for $16,627,739.70, a roughly tenfold jump in 24 hours. The emails came through Stripe, Anthropic’s payment processor, and the user said the sender address and payment link both resolved to Anthropic’s official domain, ruling out the phishing he first suspected. His bank declined two overseas charge attempts from a merchant listed as “ANTHROP” for exceeding his card’s per-transaction limit, so no money moved.
❗️ Anthropic tried to charge a Korean user who was on the free plan with zero API usage $16.6 million. A day earlier, the same invoice was $1.67 million, so it grew roughly 10x overnight.
The user says he suspected phishing at first, then found the sender and payment link were… pic.twitter.com/u9g4wjYiZ6
— International Cyber Digest (@IntCyberDigest) July 10, 2026
The details that make the episode striking are also the ones that remain unexplained. The user says he was on the free tier, had no payment card registered, and saw zero API activity on his dashboard, yet charge attempts still reached a bank card. How an unregistered card was billed has not been clarified.
Anthropic told ZDNet it is difficult to confirm specifics of individual accounts, and as of the initial reports had not publicly stated whether the invoices were a billing error. Later coverage reported the user spent four days and some 18 emails seeking confirmation the invalid invoices were cancelled, and quoted him saying the response felt inadequate given the size of the error. Anthropic’s account of what happened remains limited, so much of this rests on one user’s screenshots.
The technical cause is unconfirmed, but developers have pointed to a plausible mechanism. AI agents that call the API in loops can generate thousands of billable requests when they get stuck, a pattern known as a “retry storm,” where each retry is a separate charge. That could explain runaway costs for a heavy user, but not a free-tier account reporting no usage, which is why the case looks more like a billing-system malfunction than genuine consumption. A structural issue compounds the worry: token counts are calculated on Anthropic’s servers and cannot be independently verified by customers, unlike metered electricity or bandwidth.
Part of a Pattern
This is not an isolated complaint. Users on Reddit and GitHub have reported contradictory Anthropic billing emails and repeated small erroneous charges for months. In one case, a Claude Max subscriber was charged about $201 in extra usage credits while most of his weekly plan sat unused, and Anthropic initially declined a refund, calling it an “un-refundable technical error,” relenting only after the complaint drew more than a million views online.
The Head of Claude Code acknowledged that instance as an “overactive anti-abuse system” and marked it fixed. Separately, the auditing startup Vaudit said it found roughly $1.7 million in overcharges across $34 million of AI invoices it reviewed, much of it tied to Claude Code, and a proposed class action has challenged whether Claude Max plans deliver the usage advertised. Together these suggest recurring strain in Anthropic’s billing and support systems.
The Trust Problem
The core issue is verifiability and recourse. Because customers cannot audit token counts themselves, they must trust the provider’s meter, which makes billing accuracy and responsive support essential rather than optional.
Several reports describe Anthropic’s consumer support leaning heavily on an AI chatbot with unclear escalation to a human, and refunds arriving at the speed of viral attention rather than the speed of the error. For a company that markets itself on safety and trust, billing failures cut directly at that reputation, especially for developers weighing whether to build businesses on the Claude API.
No money ultimately changed hands in the $16.6 million case, and a single viral screenshot warrants caution until Anthropic gives a full account. But the accumulation of similar complaints, not the eye-catching figure alone, is what makes the story matter, and it points to a fix Anthropic has yet to fully deliver: transparent, promptly corrected billing with a clear path to a human.
