Detailed Analysis
A bug affecting Claude's usage limit tracking emerged shortly after the release of Opus 4.7, with users reporting that their 5-hour and weekly usage windows were resetting and clearing continuously rather than enforcing limits in the expected manner. The issue appears to be concentrated in the Claude Code CLI environment, where local storage accumulates session data that overflows during limit calculations, triggering automatic resets instead of properly pausing usage until the window expires. Symptomatic error messages such as "Claude Code failed to load usage data reset reason overflow" and "Claude usage limit reached" — accompanied by reset timers that never actually trigger — have been documented across multiple GitHub issues, including #9005, #8675, and #9048 in the anthropics/claude-code repository. The web interface, by contrast, appears to function normally during the same episodes, pointing to a client-side rather than server-side origin.
The root cause is closely tied to how Opus 4.7's release altered certain underlying parameters — notably the removal of sampling parameters — which appears to have disrupted the local usage tracking logic that the CLI relies on to manage and enforce rate limits. When the cache accumulates sufficient malformed or surplus session data, the tracking system enters a loop of continuous resets rather than stable enforcement. Practical remediation steps have been identified by the community and involve updating the Claude Code CLI to patch overflow-related bugs, clearing the usage data file stored in the hidden `~/.claude/` directory, and restarting the CLI session. For most users, these steps resolve the problem within minutes, though edge cases involving authentication state may require logging out and back in through the web interface as well.
The broader significance of this incident lies in what it reveals about the fragility of local state management in rapidly iterating AI developer tooling. Claude Code is a relatively young CLI product operating in an ecosystem where model updates — such as the transition to Opus 4.7 — can introduce subtle but disruptive incompatibilities with client-side infrastructure. The recurrence of similar overflow bugs across multiple reported episodes in late 2025 suggests that usage tracking is a persistent weak point in the CLI's architecture, likely because limit enforcement logic must coordinate between local cache state and server-side session data in ways that are sensitive to model-level changes. This pattern is not unique to Anthropic; rapid model iteration across the AI industry regularly outpaces the stability of the developer tooling built around those models.
The timing of the issue — appearing immediately after a major model drop — also underscores a recurring challenge for AI companies managing developer relations at the frontier of capability releases. When model upgrades break adjacent tooling, even temporarily, the disruption disproportionately affects power users and developers who rely on continuous CLI access for production workflows. Anthropic's GitHub issue tracker serves as the primary coordination mechanism for these incidents, and the community-driven documentation of workarounds reflects the degree to which users have taken on informal QA roles in the post-deployment phase. Whether Opus 4.7's introduction of features like Routine for scheduled sessions will eventually help stabilize usage window management — by moving limit tracking to a more structured scheduling layer — remains to be seen, but the current episode highlights the need for more robust decoupling between model versioning and client-side rate-limit enforcement.
Read original article →