When AI tools mislabel models and inflate bills
A free tool that builds personalized AI learning-path PDFs racked up a $31 bill on AWS—despite only 14 visits. The culprit wasn’t the app or its users, but an AI agent that spent a holiday weekend replaying cached context over and over.
The mismatch between visits and costs
Clew Directive, a stateless web app that generates brief PDFs after a quick questionnaire, showed 14 visits in a month via Umami. Yet AWS invoiced $31, or roughly $2.21 per visit—an implausible rate for a tool that uses Amazon Nova, a low-cost model. The numbers didn’t add up: the app neither stores user data nor runs expensive models like Sonnet.
The agent that didn’t belong
Using Amazon Q and reading the code confirmed the app itself wasn’t responsible. The IAM policy allowed only Nova models, and the repo ran Nova Lite and Nova Micro. Something else was calling models, and it left a clear fingerprint: 28 million tokens processed across just eight days, with two days accounting for 70% of the cost. The pattern—heavy cache writes, huge cache reads, and minimal input or output—pointed to an agent repeatedly loading and re-reading a large fixed context.
What really happened
The developer traced the spike to a personal project run over Memorial Day weekend. That unrelated code had cached a large prompt and kept reusing it, triggering Nova calls that the AWS bill attributed to Clew Directive. The episode underscores a growing risk: when AI agents operate outside the intended application, cloud costs can balloon even when user traffic is minimal.
Source: DEV Community. AI-assisted editorial synthesis — TechnoExpress.