Amazon Claude job billed $1.8M and blew 860% over budget

A single Amazon side project reportedly billed $1.8 million to Anthropic's Claude Sonnet before anyone at the company realised what was happening, according to Tom's Hardware summarising details first surfaced by the Financial Times from an internal Amazon presentation. The task was mundane: match author records to product listings on the retail site. It ran roughly 860% over its original budget, went undetected for about five months, and never shipped. Senior engineers reportedly called it "catastrophically expensive."

The interesting part is not the number, it is how the number happened. When traditional code has a bug it crashes and someone gets paged. When a misconfigured agent has a bug, it just keeps looping, quietly consuming tokens, and the only real signal is a bigger invoice at the end of the month. The Next Web reports that Amazon flagged two other overruns at the same meeting, roughly $541,000 on a financial auditing tool and $134,000 on a delivery acceleration system, so the author-matching bill was not a one-off but a pattern.

Why this matters if you are nowhere near Amazon's scale: the same failure mode is available to any team letting an agent retry, self-critique, or loop without a hard spend cap. AWS Bedrock itself offers batch inference at a 50% discount, prompt caching that can trim input costs by up to 90%, and Haiku models at roughly a third the price of Sonnet. Amazon reportedly did not lean on any of those on the jobs that blew up.

The honest caveats. The reporting traces back to an internal presentation and Amazon's public line is only that "as with any new technology, we're experimenting, learning and improving how we use it." What the reporting does not give you is which team owned the project, why no billing alert fired for five months, or how much Amazon has lost in aggregate across similar unshipped experiments. Take the specifics as reported, not settled.

The forward-looking read is that FinOps for agents is about to be a real product category rather than a checkbox. Whoever ships credible per-job spend caps, live token dashboards, and cheap fallback routing to Haiku or batch inference gets to sell into a market that just watched the biggest cloud on earth lose seven figures to a script that would not stop retrying.