When a Non‑Engineer Deploys with Claude Code, a Hidden Bug Makes One Day of AI Cost a Month of Server Fees

Jumpei Ueno, a senior cloud‑infrastructure engineer, inherited a project built by his company’s CFO using Claude Code, who created and deployed a SaaS product in just two days before handing it over for maintenance.

While reviewing the LLM API cost chart, Ueno noticed a single day’s spend that alone accounted for half of the month’s total API bill, far exceeding the cost of running the entire server fleet for a month.

His first hypothesis was that repeated manual testing during development had accumulated many expensive calls, as the commit history showed over twenty AI‑related commits in one day. However, detailed log analysis of the task queue, database, and request records revealed a different picture: a single batch task was executed 21 times.

The batch task consists of two steps: (1) issuing a series of requests to multiple LLMs, which generates the bulk of the cost, and (2) writing the returned results to the database. The production database had not yet received a migration that added a new column required by the code, so the write operation failed with a “column does not exist” error, causing the task to return a 500 status.

All LLM calls succeeded and were billed (each returned 200), but the failure occurred after the successful inference. The task queue treats a 500 response as a transient error and automatically retries the entire task. Because the task is not idempotent, each retry repeats all LLM calls, leading to a “retry storm” where the same successful calls are billed repeatedly.

Two root causes were identified: (1) deployment order error – code was released before the database migration, creating a deterministic failure that never resolves on its own, and (2) the task queue’s default automatic retry on 500 errors, which repeatedly re‑executes a non‑idempotent workflow.

Ueno distilled several lessons: deterministic failures must not be retried indefinitely; set explicit retry limits; ensure costly operations (e.g., billing‑related API calls) are idempotent so that repeated attempts can skip already‑completed work; always apply database migrations before deploying code that depends on new schema; and monitor API usage with budget alerts to catch anomalies early.