OpenAI says my prepaid credits were consumed, refuses to show any record

The incident began when a customer's credit balance dropped from 491.8 credits to zero within a second, despite no logged API calls. Subsequent purchases of 1,000-credit packages failed to appear, and the Codex product vanished from the account before reappearing with part of the newly bought credits already gone. The usage dashboard also stopped reporting any activity after the balance wipe, leaving the user without any visible metric of consumption.

OpenAI's support team responded to multiple inquiries with statements that they lacked visibility into the detailed account activity and could not provide itemized logs or internal ledger records. Over a period of weeks, the same phrasing was repeated in writing, indicating a systemic limitation in the support tooling rather than an isolated oversight. A formal GDPR request for the account records was met with deferrals and references to the user's own local files, prompting a regulatory complaint.

For engineers building on OpenAI's API, this lack of transparency means that reliance on the provider's dashboard for cost monitoring is insufficient. Unexpected spikes in credit consumption can occur during service outages or when large contexts are repeatedly sent, as described in the user's experience of four rapid purchases totaling $193.60 within 96 minutes. Without an official audit trail, teams must implement independent logging of token usage and enforce hard budget caps on their side.

Mitigating the risk involves adding application-level instrumentation to record request payload sizes, token counts, and timestamps, and coupling that data with alerts when consumption exceeds predefined thresholds. Developers may also consider setting up secondary budget controls, such as API gateways that enforce rate limits or abort calls once a projected cost ceiling is reached. These safeguards require engineering effort and may introduce minor latency, but they provide visibility that the provider currently does not offer.

The cost of adopting these measures is primarily development time and the overhead of maintaining custom monitoring infrastructure. The approach stops being effective only if OpenAI completely disables any usage metrics from the API response, in which case external logging would be the sole source of truth. Until the provider supplies reliable, itemized records, engineers must treat OpenAI credit consumption as an opaque external dependency.