- Anthropic acknowledged that Claude Code users are “hitting usage limits way faster than expected,” with some developers exhausting their 5-hour session quotas in under 90 minutes.
- Contributing factors include intentional peak-hour throttling affecting roughly 7% of users, the end of a promotional period that doubled off-peak usage, and suspected prompt caching bugs that may inflate token consumption by 10-20x.
- The issue affects both Pro ($20/month) and Max ($100/month) subscribers, with developers reporting that automated workflows silently deplete budgets within minutes.
- Anthropic called the issue “the top priority for the team” but has not disclosed exact usage limits for any subscription tier.
What Happened
Claude Code users began reporting abnormally fast token consumption starting around March 23, 2026, with usage quotas draining far more quickly than in previous weeks. Anthropic acknowledged the problem, stating that “people are hitting usage limits in Claude Code way faster than expected” and that the company was “actively investigating.”
One developer on the Max 5 plan ($100/month) reported consuming their entire monthly quota in a single hour of work. A Pro subscriber said their account “maxed out every Monday and resets at Saturday,” limiting usable days to approximately 12 out of 30 per month. Reports surfaced across developer forums, Hacker News, and social media throughout late March.
Why It Matters
Claude Code has become a primary development tool for a growing number of software engineers who rely on it for code generation, debugging, and refactoring throughout their workday. When usage limits are reached unexpectedly, developers lose access to a tool that has become embedded in their workflow, creating disruptions that affect project timelines and productivity. For developers paying $100 per month on the Max plan, the expectation is that the tool remains available for a full working day.
The issue is compounded by Anthropic’s deliberate opacity about exact usage limits. The company describes plan tiers only in relative terms, such as “at least five times” the baseline, without disclosing specific token counts. This makes it difficult for developers to plan their usage or diagnose whether they are hitting a bug or a legitimate limit.
Technical Details
Multiple factors contributed to the accelerated quota consumption. On March 26, Anthropic implemented intentional peak-hour throttling during weekday hours between 5 AM and 11 AM Pacific Time, causing approximately 7% of users to hit session limits they would not have encountered previously. Two days later, a promotional period that had doubled usage limits during off-peak hours from March 13 through March 27 ended, creating a jarring return to normal capacity for users accustomed to the higher allocation.
Users who reverse-engineered the Claude Code binary reported finding two independent bugs that caused prompt caching to break, silently inflating token costs by 10-20x. When the cache fails, every interaction requires sending the full conversation context to the API rather than referencing cached tokens, dramatically increasing the token count per request. The five-minute cache lifetime means that even brief breaks from coding require full context reprocessing, further accelerating consumption. Some users reported that downgrading to Claude Code version 2.1.34 improved consumption rates, suggesting the bugs were introduced in a recent update.
Who’s Affected
Both Pro and Max subscribers are affected, but developers running Claude Code in automated loops face the most severe impact. When rate-limit errors trigger unhandled retries in automated workflows, budgets can be depleted within minutes without any visible warning. Individual developers doing interactive work report that sessions that previously lasted a full workday now expire within one to two hours.
The problem is particularly acute for teams that have standardized on Claude Code for their development workflow and have no equivalent fallback tool configured. Some developers have switched to GitHub Copilot or Cursor as temporary alternatives, while others report managing consumption by batching tasks and avoiding usage during the 5 AM to 11 AM Pacific peak window when throttling is most aggressive.
What’s Next
Anthropic described the issue as “the top priority for the team” but has not provided a timeline for a fix or committed to disclosing exact usage limits. Until the caching bugs are resolved and throttling policies are clarified, developers face unpredictable access to a tool many have integrated into their daily work.
The incident highlights the risk of depending on AI development tools where the provider controls usage metering without transparency. Developers cannot effectively budget their usage when exact token limits are undisclosed and consumption rates can change without notice due to backend throttling decisions or software bugs.
