The Amazon Claude overspend reached $1.8 million on one internal project, according to the Financial Times, after a Claude Sonnet deployment ran far past its budget before Amazon caught it. The system was meant to match author information with Amazon listings, a dull data-cleanup job that became expensive once an AI model was put in the loop.
The Financial Times reported that Amazon internal reports describe several cases where AI use pushed project costs beyond expectations. In the largest disclosed case, the Claude Sonnet project came in 860% above its planned budget and was detected roughly five months after the overspending began.
Two other examples cited in the reporting were smaller but still expensive: a financial auditing tool added $541,000 in costs, and a logistics system intended to cut delivery times generated an extra $134,000 expense.
Why did Amazon’s Claude cost overrun happen?
The problem appears to be metered AI usage at scale. Models such as Claude process prompts and outputs as tokens, the small chunks of text used by large language models. If software keeps asking the model to perform repetitive work, retry tasks, or call other tools, token usage can climb quietly until the bill is no longer quiet.
That is the annoying part for engineering teams. A bad loop in conventional software might waste cheap compute or produce a pile of logs. A bad loop wrapped around a commercial AI model can turn every needless pass into a billable event.
What did Amazon say about the AI spending examples?
Amazon pushed back on the idea that these cases reflect normal AI use inside the company. In an internal presentation quoted by the Financial Times, Amazon said it is “experimenting, learning and improving” with the technology, including on cost controls. The company also said selecting “small, isolated examples” from teams learning from each other does not represent how Amazon teams use AI overall.
The numbers are small beside Amazon’s overall business. The company’s latest quarterly revenue was more than $181 billion, according to the report. Even so, the internal cases show why AI cost controls have become a real engineering issue rather than a finance-department footnote.
Amazon has already dealt with other AI workflow problems. Earlier reporting described AWS outages connected to AI coding bot mistakes, after which Amazon limited AI agents’ access instead of letting them operate with permissions comparable to the senior engineers associated with them.
The company also previously used an internal leaderboard showing which employees used AI the most, according to the reporting, but later dropped it as rising AI costs made that kind of incentive look less clever. Pushing workers to generate more model usage is easy. Proving that the extra tokens ship better software is the harder, less slogan-friendly part.
Other technology companies have run into the same tension as AI providers shift more spending toward per-token pricing and as agentic systems increase the amount of model work behind a single user request. The Financial Times report gives the phenomenon a concrete Amazon price tag: one ordinary internal task, one delayed cost alarm, and a seven-figure lesson in metered automation.
This story draws on original reporting from Tom's Hardware.
