Detailed Analysis
A Claude Max subscriber reported on Reddit that a platform-wide outage consumed the final 10% of their weekly usage allowance without any actual usage taking place, pushing their account from 90% to 100% capacity in a matter of minutes while their 5-hour usage counter remained completely unchanged. The user, who operates two separate Claude Max accounts and monitors both via the Claude Code CLI status line, was certain no genuine consumption occurred: the affected account was not actively being used during the outage, and the discrepancy between the weekly and 5-hour counters provided a clear internal contradiction that pointed to erroneous counter behavior rather than legitimate usage. After reaching Anthropic's support via the claude.ai help chat interface, the user received minimally helpful responses from the AI support agent "Fin," prompting them to escalate the issue publicly while expressing goodwill toward Anthropic's broader track record.
The incident fits within a documented pattern of usage-counter anomalies that has been accumulating across Claude Code's GitHub issue tracker since at least late September 2025. Multiple users across Pro, Max 5x, and Max 20x subscription tiers have reported weekly limits being reached after two to three days of inactivity followed by only a few hours of actual use, with some users hitting limits almost immediately after restarts or following error conditions. Temporary workarounds such as logout/login cycles have been reported to briefly clear the visual indicator, but limits consistently reappear within roughly two hours, suggesting the underlying counter state persists server-side regardless of session resets. The pattern worsened noticeably around the time of the Claude 4.5 release in late 2025, according to community reports, with no confirmed resolution from Anthropic as of that period.
The broader context involves significant structural changes Anthropic made to its usage-limit architecture beginning in August 2025, when the company introduced rolling 7-day weekly caps alongside 5-hour resets, targeting a stated goal of deterring account sharing and API reselling. While Anthropic indicated these hard limits would affect fewer than 5% of users, reports suggest misconfigured rate limiters deployed around September 29, 2025 substantially reduced effective access for a much wider population, with Pro users seeing effective weekly access drop from an estimated 40–80 hours down to 6–8 hours in some cases. The outage-triggered depletion described in this Reddit post represents a particularly sharp edge case: not only are counter bugs consuming limits during legitimate use, they are now consuming limits during periods of confirmed non-use, compounding the reliability problem.
The support experience described in the post highlights a secondary concern about Anthropic's incident-response infrastructure. When billing-level errors occur — ones that directly reduce a paid subscriber's remaining access — routing users through a generalist AI support agent that responds with "please describe in detail" after multiple detailed descriptions represents a mismatch between the severity of the issue and the resolution pathway available. For a product like Claude Code, where heavy professional users depend on predictable weekly access windows to schedule workloads, erroneous limit depletion during outages is not a minor inconvenience but a direct disruption to paid productivity capacity. The user's mention of overage coverage potentially being charged during the outage further suggests financial impact beyond mere access restriction.
The cumulative weight of these reports points to a systemic instrumentation problem in how Anthropic's usage-metering infrastructure handles error states, session interruptions, and service outages. When a rate-limiting or accounting system continues to increment usage counters during periods where the underlying service is unavailable, it suggests that the counter update and the actual inference computation are not atomically coupled — charges or limit deductions can be recorded without corresponding successful outputs. Resolving this class of bug typically requires both a server-side audit and rollback mechanism for affected accounts and a retroactive credit process, neither of which appears to have been formalized in Anthropic's public-facing support workflow as of the time of the report. The incident underscores the growing pains associated with deploying sophisticated metering systems at scale as Anthropic transitions Claude from a primarily consumer chatbot to a professional developer infrastructure platform.
Read original article →