Detailed Analysis
Claude Code users across multiple subscription tiers began experiencing unexpectedly rapid depletion of their usage limits in late March 2026, prompting Anthropic to acknowledge the problem publicly as a "top priority." Reports from affected users describe session windows designed to last several hours being exhausted within minutes: Pro subscribers found their allocations gone in as little as 10 to 15 minutes, Max 5 users saw one-to-two-hour windows drain in place of the expected five, and at least some Max 20x subscribers watched their usage leap from 21% to 100% on a single prompt. The severity and breadth of these reports across tiers indicate a systemic issue rather than isolated edge cases.
Three distinct but overlapping causes have been identified. First, Anthropic began applying more aggressive usage throttling during weekday peak hours — 5am to 11am Pacific Time — starting in late March 2026, a policy change that affected roughly 7% of users but went without formal public announcement. Second, a temporary promotion that had doubled usage limits outside peak hours expired on March 28, 2026, silently reverting subscribers to standard allocations and amplifying the perception of sudden scarcity. Third, and perhaps most technically significant, Claude Code was found to contain bugs that cause prompt cache failures, silently inflating token costs by a factor of 10 to 20 times. Some users reported measurable relief by downgrading to version 2.1.34 of the tool, suggesting the caching defect was introduced in a more recent release.
The communication dimension of this episode is notable. Rather than proactive disclosure through official channels, information about the throttling changes and promotion expiration surfaced primarily through posts by Anthropic engineers on social media. This gap between policy implementation and transparent user communication left subscribers confused about why their workflow economics had changed without warning. The combination of an undisclosed policy shift, a quietly expiring promotion, and a software bug created compounding effects that users had no framework to diagnose or anticipate.
The incident connects to a broader tension in the AI tooling market between the rapidly expanding computational demands of agentic coding tools and the subscription pricing models designed to support them. Claude Code, as an agentic coding assistant, is architecturally prone to high token consumption because it operates in extended multi-turn sessions with large context windows and iterative file manipulation. The prompt caching mechanism exists precisely to contain those costs, making its failure particularly disruptive. As AI coding assistants mature from novelty features into core development infrastructure, the reliability and predictability of usage economics become critical factors in enterprise and professional adoption.
Anthropic's situation reflects a challenge facing the AI industry broadly: scaling infrastructure and pricing models fast enough to meet demand from power users who are pushing tools to their limits in ways that outpace original consumption projections. The fact that a caching bug could go undetected long enough to affect a significant user population points to gaps in observability tooling for agentic workloads. For Anthropic, which has positioned Claude Code as a flagship product in the competitive AI coding assistant space — alongside offerings from OpenAI, Google, and startups like Cursor — restoring user trust will require both technical remediation and more transparent communication practices around pricing and policy changes.
Read original article →