Anthropic and OpenAI customers overcharged by $1.7M in billing errors, startup audit finds

Anthropic and OpenAI are helping companies automate code, customer support, research, and internal workflows. Now some of those same companies are finding they may have been paying too much for the privilege.

A startup called Vaudit says it found roughly $1.7 million in billing overcharges after reviewing $34 million in AI invoices sent to 60 companies between March and June, most of which were tied to Anthropic’s Claude Code. The audits covered bills linked to Anthropic and OpenAI usage, and the findings point to a growing problem for enterprise buyers racing into AI without a clear view of what they are actually being charged for.

Vaudit CEO Michael Hahn told The Information that common issues included customers being billed for newer, more expensive models when they were actually using older, cheaper ones. In other cases, customers were charged when AI agents or chatbots failed to complete requests or returned error messages. Hahn said Vaudit also found cases where failed agents kept retrying tasks in the background, quietly running up bills in what’s known as a retry storm.

“Between March and June, Vaudit audited bills sent to 60 companies totaling $34 million, mostly for usage of Anthropic’s Claude Code, and found about $1.7 million in mistaken overcharges,” The Information reported, citing Hahn. 

“What we are observing is that enterprise AI billing has become increasingly opaque,” Hahn said.

The Hidden Cost of AI: Audit Finds $1.7M in Anthropic and OpenAI Billing Mistakes

The claim lands at a moment when enterprise spending on generative AI is climbing fast, and finance teams are under pressure to make sense of usage bills that can change by the hour. AI pricing already comes with enough moving parts: model tiers, token usage, API calls, retries, cloud markups, and different billing paths depending on whether customers buy directly from model providers or through large cloud platforms. Vaudit’s pitch is that many companies lack the internal tools to verify invoice accuracy. Vaudit counts Panasonic, HP, and Honda among its customers, Hahn said.

Anthropic pushed back on parts of the report. A spokesperson said the company does not charge customers for incomplete requests or error messages, does not route customer requests to older models, and does not see signs that overbilling is a widespread issue. OpenAI said it has no evidence that those issues are happening among its customers.

That leaves an important caveat in the middle of the story: not every billing issue described by Vaudit necessarily points to a direct error by Anthropic or OpenAI. Hahn said some of the audited customers accessed those models through major cloud providers, which adds another layer between usage and invoicing. That matters, since charges can pass through Amazon, Google, or Microsoft before they ever land on a customer’s finance dashboard.

Still, Vaudit says the money was real. Hahn told The Information that after Vaudit and its customers challenged the invoices, about 80% of the disputed overcharges were credited back by Amazon, Google, Microsoft, Anthropic, and OpenAI. He described those companies as “incredibly cooperative” when billing issues were flagged.

The bigger takeaway may be less about any one disputed invoice and more about the shape of the enterprise AI market itself. Companies have spent the past two years rushing to deploy copilots, internal assistants, code agents, and customer support bots, often with only a rough sense of what production usage will cost at scale. A team can start with a promising pilot and end up with a monthly bill that is hard to audit, forecast, or challenge.

As AI spending rises, so do questions about what enterprises are actually paying for

That billing confusion is starting to show up elsewhere. Anthropic is already facing scrutiny over pricing from another direction. Last week, a customer filed a proposed class action lawsuit accusing the company of misleading users about the value of its higher-priced subscription tiers. At the same time, many businesses that were once eager to spend freely on premium frontier models are starting to ask whether they really need them. Some are moving parts of their workloads to cheaper models or open-source alternatives to rein in costs.

Vaudit is trying to turn that anxiety into a business. The 30-person startup, founded in 2023, sells software that customers install within their AI environment, typically via a software development kit, to capture raw usage data and compare it with invoices and billing records. Hahn, a former Oracle director, originally built the company to audit bills tied to logistics, shipping, advertising, and cloud services before launching its AI billing product earlier this year.

The company says it charges customers 1% of the value of the bills it reviews, plus 30% of any refunds recovered after an audit. That pricing model only works if there is money to find. Vaudit says there is.

For enterprises betting big on AI, the message is uncomfortable but simple: the hard part may no longer be getting employees to use these tools. It may be figuring out whether the invoice at the end of the month makes any sense.