Amazon's $1.8 million Claude mistake highlights the escalating costs associated with AI.

      Amazon has discovered that artificial intelligence can escalate minor errors into costly ones totaling seven figures. Internal metrics reveal that several projects have overshot their budgets dramatically, including one that incurred $1.8 million in losses before anyone noticed. Senior engineers labeled the findings as "catastrophically expensive."

      These situations were presented in an internal meeting and were first reported by the Financial Times. The most significant overrun involved using Anthropic’s Claude to compare author records with product listings in Amazon’s store. This particular project exceeded its budget by 860 percent, went unnoticed for five months, and ultimately failed in deployment.

      Tom’s Hardware also highlighted two smaller mistakes alongside this major blunder. A financial auditing tool overspent by approximately $541,000, while a system designed to expedite deliveries added an extra $134,000. Each of these errors was the type that previously incurred minimal costs.

      How a minor mistake escalates into a large expense

      The issue lies in the billing structure. When a developer writes faulty code, it typically results in an error and a crash. However, when a model is involved, a misconfiguration can persist without interruption, racking up costs quietly. For instance, a retry loop that sends the same prompt multiple times or a job targeting the entire catalog instead of a sample won’t crash the system; it will generate an invoice instead.

      This invoice is issued on a monthly basis, rather than appearing in a build log. This delay explains how a misconfigured job can run for five months before it is linked to a financial figure. The shift to per-token pricing exacerbates this issue. AI agents, which utilize significantly more tokens than a chatbot, only intensify the problem.

      This problem isn’t exclusive to Amazon. It mirrors the situation that led to a $1.3 million bill from OpenAI for another developer and explains why costs for enterprise AI continue to rise, even as the per-token price decreases.

      The company providing the solution

      Here’s the irony: Amazon overspent on AI while its own cloud services offer the very tools needed to prevent such overspending. AWS’s Bedrock provides batch inference at half the cost, a discounted “Flex” tier, prompt caching at a tenth of the input rate, and prompt routing that directs simpler tasks to less expensive models.

      Cheaper models were available as well. Anthropic’s Haiku costs only one-third of Amazon's chosen Sonnet model. The overspending was a result of defaulting to the most advanced model without implementing safeguards, which is precisely what both Amazon and Anthropic advise others to avoid.

      Amazon’s response

      Amazon downplayed the situation, stating, “Like any new technology, we are experimenting, learning, and improving our usage.” They emphasized that these isolated incidents do not represent the overall operations of their teams, noting that they involved a small number of groups within a corporate workforce of approximately 300,000.

      The financial implications support this nonchalant stance. Amazon’s quarterly revenue exceeds $180 billion, meaning the overspent budgets don't even account for 0.1 percent of a single month’s sales. AWS cost expert Corey Quinn commented without sympathy, suggesting that “receiving an unexpected bill with two commas must be quite challenging for you.”

      However, behind the scenes, the response has been more decisive. Amazon has eliminated an internal leaderboard that ranked employees by AI usage after staff manipulated it by inflating their token consumption. Its engineers are now developing automated safeguards to limit project expenses before the bill arrives.

      This is the fundamental lesson, focused on governance rather than finances. For Amazon, a loss of $1.8 million is negligible. For the multitude of businesses now relying on AI to reduce costs, a runaway job that remains undetected for five months could mean the difference between staying within budget and facing a crisis.