Amazon reportedly burned $1.8 million on a single Claude project

Amazon accidentally burned $1.8 million on one Claude task. As first reported by the Financial Times on July 30, 2026, an internal Amazon project used Claude to match author records to product listings on the retail site. The task ran 860% over budget, quietly consumed about $1.8 million over five months before anyone noticed, and never launched. Senior engineers called it “catastrophically expensive” in internal reviews.

Not an isolated incident

The author-matching project had company. A financial auditing tool went about $541,000 over budget, and a delivery-optimization system added $134,000 in overruns. These are three separate teams inside the same company losing track of AI costs in parallel.

The leaderboard that made it worse

Culture played its part. Amazon pushed for 80%+ of developers to use AI weekly and ran an internal leaderboard ranking staff by AI usage. Employees gamed it the way people game every metric: by inflating token consumption. The company eventually scrapped the leaderboard, but by then the incentive had done its work, rewarding usage volume rather than usage value.

AWS cost expert Corey Quinn could not resist the irony of the world’s largest cloud provider getting surprised by a cloud bill: “getting a bill with two commas that you weren’t expecting must be so hard for you.”

Why token billing breaks budgets

The bigger story is structural. Token-based billing means costs scale with model chatter, not with results. A verbose agent that retries, elaborates, and loops burns budget whether or not it produces anything useful. Traditional software costs are mostly fixed per deployment; agent costs are variable per token, and nothing in the bill distinguishes productive tokens from wasted ones.

That is what makes the Amazon story broadly relevant rather than merely embarrassing. If the company that runs the world’s largest cloud can lose track of an AI bill for five months, most organizations are flying blinder than they think. The practical takeaway fits in one line: go check your own token dashboard, and make sure someone owns the number on it.