Amazon Lost $1.8 Million on a Simple Task Assigned to Claude

An internal meeting at Amazon was supposed to be a routine review of AI spending. Instead, it revealed that one project using Anthropic's Claude Sonnet model had burned through $1.8 million on a task that should have cost a fraction of that amount. The company only found out five months after spending had started spiraling out of control.

The project in question was meant to solve a fairly mundane problem: matching author records to the corresponding product listings in Amazon's online store. In a traditional software development model, this task would have cost relatively little, even if the code contained bugs. With a language model billed per token, though, every inefficient loop or poorly designed process turns into a real bill.

How the Budget Overrun Happened

According to Financial Times reporting cited by Tom's Hardware, at Claude Sonnet's input rates, $1.8 million corresponds to roughly 600 billion processed tokens. Such enormous usage points to a serious design flaw in the process itself, either an infinite loop or repeated execution of the same queries without a control mechanism to stop the process once it passed a reasonable limit.

Notably, the project never actually made it to production. Months of work and millions of dollars spent on tokens never translated into any deployed solution. For comparison, Amazon's quarterly revenue reaches roughly $180 billion, so the loss itself doesn't threaten the company's finances, but it points to a pattern that is starting to worry the controlling department.

Not an Isolated Case

The Claude Sonnet mishap wasn't a one-off. In the same internal presentation, Amazon disclosed two more instances of AI project budget overruns. A financial audit tool generated an extra $541,000 in costs, and a system meant to speed up logistics processes went $134,000 over plan, though in that case the error was caught much faster, within two weeks.

It's very hard to understand what anything related to AI actually costs - unnamed senior Amazon employee

Asked about the matter, an Amazon spokesperson offered only a general statement that these cases involved just a handful of teams out of the company's hundreds of thousands of employees.

As with any new technology, we're experimenting, learning, and improving how we use it - Amazon spokesperson

A Ranking That Spiraled Out of Control

The episode also revealed an interesting detail about Amazon's internal culture. The company ran an internal ranking of employees based on how heavily they used AI tools, treated somewhat like a step counter in a fitness app. The problem was that some teams started artificially inflating their token usage to rank higher, which drove costs up even further. Amazon dropped the practice once costs began spiraling out of control.

This isn't the first high-profile case of a costly oversight when using Anthropic's models at a large organization. Previously, an unnamed company was reported to have spent as much as $500 million on Claude in a single month because it had set no usage limits on employee licenses. In both cases, the common denominator is a lack of cost oversight mechanisms, not the quality of the model itself.

What This Means for AI Deployment

Amazon's story lands at a moment when more and more organizations worldwide, including in Poland, are grappling with rising AI token bills despite falling per-token unit prices. The reason is a sharp rise in usage, driven especially by agentic systems that make far more queries than a single user typing in a chat window. Experts advise companies to set hard token limits per employee or project, route simple queries to cheaper models, and monitor usage continuously rather than reviewing bills after the fact.

For IT and finance departments, Amazon's case is a practical lesson: with token-based billing, a small bug in a loop or a missing stop mechanism can cost as much as a sizable engineering team does in a year. AI cost governance, once treated as bureaucratic overhead, is becoming as important as the choice of model itself.