Detailed Analysis
A software engineer and daily Claude Code user has surfaced a series of billing and session-limit anomalies that raise questions about the reliability of Anthropic's usage tracking infrastructure. The user reports two distinct categories of problematic behavior: a "ghost exhaustion" event in which a session limit appeared to re-trigger almost instantly after a reset window had fully elapsed — with over 12 hours having passed and no intervening usage recorded — and a credit disappearance incident in which more than $3 of prepaid token balance vanished overnight following a $40–50 extra-token purchase, despite no documented consumption. The user, who describes relying on Claude Code as a core part of their professional development workflow, frames the experience as either a significant software defect or something more troubling, and expresses frustration at being unable to escalate the issue beyond automated support responses.
The session-limit anomaly described is technically notable because it suggests a potential mismatch between the client-side display of reset timing and the server-side state that actually governs access. If a limit reset timestamp is shown to a user but the underlying counter has not been cleared, or if usage from a prior session is being attributed to a subsequent one due to incorrect windowing logic, the practical result would be exactly what the user describes: a near-instantaneous re-exhaustion with no apparent cause. The credit disappearance issue is a separate but compounding concern, pointing to possible inconsistencies in how prepaid balances are reconciled against usage ledgers — a problem that, if systemic, would disproportionately affect power users who purchase additional token allocations.
The broader significance lies in what these failures mean for professional adoption of AI coding tools. Claude Code competes in a rapidly maturing market alongside GitHub Copilot, Cursor, and other developer-focused AI environments, all of which are being evaluated not just on capability but on billing predictability and support reliability. For a software engineer trying to plan sprints or estimate costs, opaque or inconsistent metering is a material workflow problem, not merely an inconvenience. The user's inability to reach a human support agent capable of reviewing raw account logs reflects a support infrastructure that has not scaled commensurately with the technical complexity of the product's billing model.
This kind of user report also illuminates a structural tension in the current generation of consumption-based AI products. Token limits, rolling windows, and prepaid credit systems introduce accounting complexity that traditional SaaS billing does not carry, and bugs in these systems are difficult for end users to diagnose or dispute because they lack access to the underlying logs. Anthropic, like other frontier AI providers, is navigating the challenge of building enterprise-grade billing reliability on top of infrastructure that was designed primarily for research and product experimentation. As Claude Code moves toward broader commercial adoption, incidents like this — whether isolated bugs or symptoms of a deeper architectural issue — represent a category of risk that could erode trust among the professional developer audience the product most needs to retain.
Read original article →