AI cost overrun: Amazon’s 860% and AI token costs

What Amazon’s AI cost overruns reportedly were

The largest reported overrun is about $1.8 million on a failed Claude Sonnet deployment, 860% over its budget and found about five months after the problem began. The job was meant to match author details with listings on Amazon. The reporting describes several internal reports behind it.

Sonnet is a model from Anthropic, the maker of Claude. Two smaller cases appear in the same report: a project building a financial auditing tool ran up $541,000 in additional cost, and a system meant to shorten delivery times in the logistics network added $134,000. As plain arithmetic that is not in the report, an 860% increase means spend reached 960% of the budget, or 9.6 times the amount allocated.

Amazon's side is reported too. An internal presentation, as quoted in that reporting, says the company is experimenting, learning and improving how it uses AI, including how it cuts cost. It adds that picking small, isolated examples of teams learning from each other, and presenting them as normal, does not reflect how Amazon teams use AI.

The reporting sets the amounts against Amazon's quarterly revenue, which it puts at more than $181 billion, and concludes that the excess spending is under 0.1% of a single month's revenue. That comparison is the outlet's, not Amazon's.

Causes of cost overrun: AI token costs billed on every retry

AI token costs turn each repeated attempt into a charge, because providers bill by the amount of text a model handles instead of a flat fee. A token is the small unit of text a model reads or writes, and per-token pricing multiplies the count of those units by a rate, so a call that runs again is counted again.

The reporting says the cost of mistakes changed. Errors that were once "trivially cheap" became "catastrophically expensive," especially as token spending rose sharply once AI agents were deployed. An agent here means a program that works through a task by calling a model over and over, so a single task can produce a large volume of billable text.

The billing model moved at the same time. As agents took over and providers switched from subscriptions to per-token pricing, costs grew large enough that some companies are using up their annual budgets in weeks.

Take a team that points an agent at a large set of product listings. A misconfigured run still executes and is still charged by volume. This is an illustration, not a description of the Amazon job.

AI cost control and governance: a five-month gap with no stated cause

The reporting gives the lag, about five months from the start of the problem to detection, and gives no reason for it. Whether alerts existed, who watched the spend, and why nothing escalated earlier are all left open. What the same article does add are three neighboring facts about measurement and permissions.

First, Amazon used to keep an internal leaderboard of which employees used AI the most. It has since dropped it, because spiraling AI costs made the company reconsider the policy. The source gives no date for the change.

Second, the Uber CTO is reported as seeing no link between pushing people to use AI, which the report calls tokenmaxxing, and shipping successful products. Third, AWS reported several outages earlier in 2026 that were driven by mistakes from AI coding bots, and the fix was to limit what AI agents can access, instead of giving them the same permissions as the senior engineers they are tied to. For a company setting up AI cost control and governance, these three items point at the same thing: what gets measured and what an agent is allowed to touch.

Why AI spending can strain companies other than Amazon

The report says per-token pricing and agents have pushed some companies through their annual budgets in weeks, and calls that unsustainable for most companies other than giants such as Amazon and Microsoft. The wording covers most other companies, not a named size band, and it gives no figures for any company besides Amazon.

The 0.1% comparison reads differently from the other side. For Amazon, the reported excess is a small share of one month's revenue, so the outlet treats it as affordable. A company without that revenue cannot assume the same sum is small, and the report offers no numbers to say how large it would be for them.

For a team that budgets AI use as an annual line, the practical question is how fast that line can be spent. Per-token billing ties the bill to usage volume, and agents raise the volume. A flat subscription fixes the monthly charge up front, which per-token billing does not.

FinOps for AI: AI cost management with fewer tokens

FinOps, short for financial operations, is the cloud-era practice of making the teams that spend money see and own that spending, and several publishers now propose applying it to AI. The Amazon reporting names no fix, so the FinOps practices are other publishers' proposals, not Amazon's. This section takes that direction from their headlines and section titles.

InfoWorld argues that the playbooks which brought cloud bills under control can also rein in generative AI costs. FinOps Weekly gives the same frame to a newsletter issue on what chief information officers learned from AI overruns. ITSM.tools puts model routing at the center, meaning a rule that chooses which model handles each request, and that is the piece closest to cutting tokens per task.

Ownership is the other recurring theme. Progressive Robot has a section on deciding who owns the number, and Luma Commons asks which team should own AI cost governance. Portal26 describes finance teams as lacking sight of token spend.

Two pieces on this site cover related ground. Claude Code Token Usage: 33K vs OpenCode's 7K looks at token use, and Claude Opus 4.6 vs 4.7 vs 4.8 pricing compared compares model prices.

Summary: the reported figures and the pattern beyond Amazon

The dollar figures come from press reporting, and Amazon disputes the framing, so they are best not treated as settled. The structure behind them does not depend on those numbers. Failure is billed on each attempt under per-token pricing, agents raised token spending, and the Amazon case shows a detection lag of about five months.

Where those three conditions meet in another company, the same pattern can form there. The report names no remedy, so the useful step is a check rather than a prediction. Whether the monitoring a company already has covers per-token spend is the question this reporting leaves open.