Amazon admits it accidentally shelled out $1.8 million for Claude to finish its menial coding tasks

Amazon revealed that its Claude Sonnet project, intended for streamlining menial coding tasks, ended up costing $1.8 million – an 860% increase over the original budget. This financial misstep went unnoticed internally for nearly five months, sparking questions about the company's internal tracking of AI expenditures. The overshoot wasn't an isolated incident, as a separate project to develop a financial auditing tool also exceeded its budget by $541,000.

Another initiative aimed at enhancing Amazon's logistics efficiency incurred an additional $134,000 in unanticipated costs. Experts note that as AI agents gain independence and token-based pricing models replace older subscription plans, the potential for overspending escalates rapidly. Amazon defended the incidents, stating they are part of ongoing experimentation and cost-efficiency improvements.

However, the company's quarterly revenue, exceeding $181 billion, means the disclosed overspending represents less than 0.1% of one month's earnings. The pattern of AI-driven cost overruns follows earlier issues at AWS, where uncontrolled coding bots caused several unexpected service disruptions. Amazon has since limited AI agent permissions and discontinued an internal leaderboard for tracking AI usage.

Despite skepticism from other industry players regarding the tangible business value of AI adoption, Amazon continues to experiment with AI-driven coding tools, citing the necessity of such investments.