$1.8 million went to an Amazon project using Anthropic's Claude Sonnet to match author details with product listings, and the overrun was spotted only after five months.
Internal reviews found the spending was not isolated: AI work also racked up $541,000 on financial auditing tools and $134,000 on delivery-speed improvements.
Amazon employees told the Financial Times AI costs were hard to track, with senior engineers warning that coding errors once cheap to fix had become "catastrophically expensive."
The overruns reflect a wider pullback from the earlier "tokenmaxxing" push that encouraged heavy AI use; Amazon shut an internal AI-usage leaderboard in May as token-based billing drove costs higher.
The episode undercuts Amazon's broader cost-cutting message after years of layoffs, including a 16,000-job round partly justified by AI-driven efficiency gains.