Anthropic refunded disputed Claude Code charges this month, then its billing system tried to charge a free user $16.6 million for usage he says never existed.
You trust that a bill for machine time reflects the machine time you actually used. Anthropic spent this week proving that trust still has to be checked.
Between March and June, auditing startup Vaudit reviewed $34 million in AI invoices billed to 60 companies, much of it tied to Anthropic's Claude Code. According to The Information, Vaudit found about $1.7 million in mistaken overcharges. The errors weren't one tidy bug. Customers were billed premium rates while requests were allegedly served by cheaper, older models. Some were charged for prompts that returned no output. Vaudit also flagged retry storms, where an autonomous coding agent hits a failed request, tries again automatically, and keeps adding charges with every attempt.
Anthropic pushed back. A company spokesperson told reporters it doesn't charge for incomplete requests, doesn't silently route customers to older models, and sees no evidence that the problem is widespread. Fair enough. But the denial sits awkwardly next to the refund number. Amazon, Google, Microsoft, Anthropic and OpenAI together refunded roughly 80% of the disputed charges.
You don't credit back four-fifths of a bill you believe was right.
Frankly, that's the more damning figure in the whole story. Not the $1.7 million. The 80%.
The Invoice That Shouldn't Exist
Days later, a second and stranger incident surfaced. A South Korean developer on Anthropic's free tier, with no payment card on file and zero API usage on his account, started receiving invoices through Stripe, Anthropic's payment processor. The first arrived on July 7 at 10:20pm for $1.67 million. A day later, on July 8 at 11:42pm, the same account was billed $16.6 million. A tenfold jump in twenty-four hours.
The charges didn't stop at the invoice. Screenshots the user shared, and later reported by IBTimes UK, show his Kookmin check card fielding two overseas charge attempts on July 8. Both were declined for exceeding the card's per-transaction limit. Both were routed through a US merchant listed simply as ANTHROP. He spent four days working through support tickets before anyone at the company gave him a straight answer.
Anthropic eventually blamed an incorrectly configured auto-reload setting, disabled it, and told the user in writing that "our payment processor attempted a charge at the invalid amount and it was declined" and that nothing had actually been collected. The company also said the incident wasn't the result of unauthorized access.
No money changed hands. But the account still generated a sixteen-million-dollar invoice for work that never happened. Four days. That's how long it took, plus a public writeup, before anyone at Anthropic said so plainly.
What Enterprises Should Demand
For anyone running agents on Claude or putting the API under a product, the point isn't that Anthropic is uniquely careless. The Vaudit findings also pulled Amazon, Google and Microsoft into the refund pool. The point is sharper than that: usage-based pricing for autonomous agents still lacks the basic protections customers expect from financial software. Nobody appears to stop a runaway agent at the first strange spike. A retry storm can run up thousands of dollars before a finance team sees the bill. A misconfigured reload setting did worse. It generated an eight-figure invoice, and the customer spotted it before the company did.
Here's the thing enterprises buying into agentic workloads should ask for. Hard spending caps that trigger automatically. Real-time usage dashboards that match the invoice line for line. A written commitment on how fast a disputed charge gets reversed, not a soft promise to investigate once a support ticket climbs the ladder.
Anthropic's credits came fast once the story went public, and that counts for something. It isn't the same as building the guardrail that would have stopped the story from happening in the first place. If companies are going to let coding agents act on their behalf, they need billing systems that can recognize when the machine has lost the plot.
Also read: TSMC's June Sales Jump 68 Percent as AI Chip Demand Keeps Climbing • China's Export Data Will Show How Much of Its Growth Now Runs on AI Chips • JPMorgan Cuts MiniMax's Price Target a Second Time as Zhipu Pulls Ahead
