Amazon reportedly spent $1.8 million on an internal Claude Sonnet deployment designed to match author details with ecommerce listings, with the failed project running 860% over budget and remaining undetected for five months.
TL;DR
- The author-matching project reportedly failed to launch after costing $1.8 million and exceeding its budget by 860%.
- Two other AI projects reportedly generated $541,000 and $134,000 in unexpected spending.
- Amazon says the cases are isolated, while engineers are working on automated cost guardrails.
Amazon’s Claude Sonnet Project Reportedly Burned Through Its Budget
According to a Financial Times report, Amazon senior engineers told colleagues that moving some work from conventional programming to AI models had created unplanned expenses. The biggest disclosed case involved Claude Sonnet, which was used to connect author information with listings on Amazon’s ecommerce platform.
The deployment reportedly cost $1.8 million, ran 860% over budget and took five months to detect. It also failed to launch, turning what was intended to be an automation project into a costly internal lesson.
The report did not identify which Claude Sonnet version was used, how many tokens the system consumed or the project’s original approved budget. As such, the incident cannot be attributed to a particular Sonnet release and an estimated token count cannot be calculated.
Amazon Staff Flagged More Unexpected AI Costs
The author-matching tool was reportedly not the only case. Amazon allegedly incurred around $541,000 in unexpected costs while developing a financial auditing tool, while an AI system intended to improve delivery speeds generated another $134,000 in accidental spending that took more than two weeks to identify.
Staff reportedly described errors that had once been “trivially cheap” in traditional systems as “catastrophically expensive” when AI models and agents repeatedly consumed paid tokens. One senior Amazon employee reportedly said that determining the cost of AI-related work remained difficult.
Amazon pushed back on the broader characterization. The company said it was “experimenting, learning and improving” how it uses new technology, including efforts to make deployments more cost-efficient, and argued that a handful of learning cases did not represent normal AI use across Amazon.
Topics For More Insights
Automated Guardrails Could Become The Main Fix
Amazon engineers are reportedly developing automated controls to prevent similar overruns. Such controls could include model-level budgets, usage monitoring, alerts and limits that block additional inference requests once a project reaches a predefined threshold.
AWS has already published a reference architecture for proactive Amazon Bedrock cost management. Its design tracks input and output token usage through CloudWatch, compares consumption against limits stored in DynamoDB and can deny a model request when the configured budget has been exceeded.
The issue arrives as Amazon rapidly expands its AI business. The company reported second-quarter 2026 net sales of $200.6 billion, while AWS sales rose 37% to $42.2 billion. Amazon also said its trailing 12-month free cash flow fell to an outflow of $7.6 billion, primarily because property and equipment spending increased by $66.1 billion, largely reflecting AI investment.
Amazon CEO Andy Jassy said, “our AI and Chips businesses each eclipsed run rates of more than $25 billion.” The contrast is notable: Amazon is generating substantial AI growth, but the Claude Sonnet incident shows that even one of the world’s largest cloud operators can struggle to connect AI usage with project-level cost controls.
