Typically, AI services like Claude API and OpenAI API charge based on actual usage, but recently a bizarre incident occurred: a user abroad shared that the Anthropic dashboard showed $0 in charges, yet they received an invoice notification for a staggering $16.62 million.
Initially thought it was a phishing email or fake notification, but after checking, they confirmed the email was indeed sent from the Stripe billing system used by Anthropic, with both the payment link and sender domain pointing to Anthropic’s official website. After the incident went public, it also sparked heated discussions on Reddit, with some suspecting it might be related to AI agent infinite loops, while others believed it was likely a billing system error—after all, even a company wouldn’t likely spend tens of millions of dollars.
Anthropic bills Claude users $16.62 million! But backend usage fees show $0
This incident was caused by a Korean developer remy_notes Shared on Threads. From his publicly shared screenshot of an Anthropic payment failure notice, the system attempted to charge $16,627,739 USD, equivalent to approximately NT$534 million, which is well beyond what a typical individual developer could generate in API bills.
Image source: remy_notes
Even stranger, after he logged into the Anthropic Console to check, the backend showed $0.00 in API usage fees, with no token usage or consumption records corresponding to this invoice. In other words, the billing data he saw didn’t match the payment notification sent by Stripe at all.
Image source: remy_notes
According to ZDNet Korea According to reports, this remy_notes normally uses a Mac mini to run multiple AI automation systems, including 8 AI agents, automated trading agents, work agents, and other automation scripts.
So when he saw the bill, he initially suspected that one of the automated processes might have gone out of control, constantly calling the Anthropic API and causing the costs to skyrocket.
However, after subsequently checking the bots, agents, proxy servers, Codex bots, and Cron jobs on the machine one by one, he claimed he couldn’t find an Anthropic API Key that could generate paid usage. Although there were Anthropic-related credentials in the system, he judged those to be just session tokens for login authentication, not keys for calling paid APIs.
If it were an Anthropic billing system error, it would normally only happen once. What makes this even more bizarre is that it appears to not be a single billing mistake—He also mentioned that bank records show Anthropic had already issued another abnormal billing transaction the day before.
According to what he shared, the first payment failure notification appeared on the evening of July 7, for an amount of $1,669,875.90 USD, approximately NT$53.66 million. By the evening of July 8, the second payment request jumped to $16,627,739 USD, almost 10 times the amount from the previous day.
A Reddit user suspected it might be a phishing email impersonating Anthropic, given that the amount clearly didn’t look like a normal bill. However, after checking the sender address and payment links, he confirmed they were both using Anthropic’s official domain. He even contacted his bank, which confirmed the charge was a legitimate automated merchant payment request from Anthropic, not simply a fake invoice.
However, both transactions ultimately showed payment failures and did not actually deduct from the account. To prevent the system from continuing to attempt charges, he has disabled the registered credit card and canceled his Claude Max subscription, while also contacting Anthropic customer support to request invoice cancellation and investigate whether his account has been subject to unauthorized access.
So developers using AI APIs must manage costs carefully. They shouldn’t just plug in an API key and let agents run unrestricted. A more prudent approach is to set up safeguards at three levels: the agent, the API platform, and the payment method. If one layer fails, the other two can still minimize losses.
On the code side, you should at least set limits on the maximum number of loops per task, token usage, concurrent execution count, and timeout duration. When an agent encounters the same error multiple times in a row, it should also immediately stop execution and notify the user, rather than continuously retrying infinitely.
When testing new agent workflows, you can first restrict usage to low-cost models and set a spending cap per test run, helping you avoid unexpected high bills.
Source: KOCPC Chinese