Amazon spent five months not noticing it had blown $1.8 million on a task that should have cost pocket change.
The Bill Nobody Caught
The project sounds almost comically mundane: match author names to book listings on Amazon’s retail site. A data-hygiene job, the kind an engineer could once have scripted for free. Amazon handed it to Claude Sonnet instead. The meter ran for five months before anyone looked at it.
The final number landed 860% over budget. The project never shipped. Senior engineers presented the numbers at an internal meeting on July 28. They reportedly called the result “catastrophically expensive” — a phrase that’s since become the unofficial epitaph for Amazon’s AI cost-control failures this year.
It wasn’t alone. A financial auditing tool ran $541,000 over its allotment. A logistics system meant to shave time off deliveries added another $134,000. Combined, the overruns hit roughly $2.5 million, according to internal documents obtained by gHacks Tech News.
Why the Math Broke
The root cause isn’t sloppy engineering. It’s a pricing model nobody fully priced in. Anthropic and OpenAI both bill by the token now, not by the seat. A bug that once cost a few wasted CPU cycles can multiply into six figures the moment it triggers an agent into a retry loop or a runaway generation spree. Amazon’s engineers made that exact point in the meeting: mistakes that used to be “trivially cheap” turned “catastrophically expensive” once AI models were doing the work.
That math isn’t unique to internal enterprise tools, either. Claude Code’s own token-based pricing shows how fast a single looping session can eat through a daily budget, which is why more teams now set hard spend caps instead of trusting the sticker price.
The Leaderboard Problem
There’s a cultural wrinkle underneath the technical one. Amazon pushed a goal of getting more than 80% of its developers using AI tools weekly. It built an internal leaderboard, KiroRank, to track who used AI the most. Employees gamed it. They assigned agents to pointless busywork just to climb the rankings. Staff called the practice tokenmaxxing, and Amazon quietly killed the leaderboard at the end of May. Senior VP Dave Treadwell’s message to staff was blunt: “Please don’t use AI just for the sake of using AI.”
It’s the inverse of how the smartest agents budget compute like a miser instead of burning through it for show.
Amazon isn’t alone in learning this the hard way. Meta ran a nearly identical internal tracker nicknamed “Claudeonomics,” built on the same logic of rewarding consumption over output. Uber’s CTO admitted in April that the company had already burned through its entire 2026 Claude Code budget.
The Real Lesson
None of this is really about Claude. It’s about what happens when a company spends two years telling employees to use more AI, then discovers “more” was never a strategy. Amazon is now building automated spending guardrails — the cost-tracking infrastructure that arguably should have existed before the mandate did.
$2.5 million barely registers against Amazon’s $200 billion AI infrastructure budget for the year. But it’s a preview of what happens at companies with thinner margins and no one senior enough to notice a runaway bill for five months.
Related: AI-Washing Is Reshaping Tech Layoffs in 2026 — The Numbers Tell a Different Story
