Claude Code Users Exhausting Quotas Far Faster Than Expected
Summary
- • Anthropic confirms Claude Code users hitting usage limits far faster than expected
- • Multiple factors converge: quota cuts, expired promotion, and suspected caching bugs
- • Bugs may silently inflate token costs by 10-20x by breaking prompt cache
- • Automated workflows especially vulnerable as rate-limit errors trigger silent retries
Details
Anthropic acknowledges crisis publicly
Anthropic confirmed users are hitting limits 'way faster than expected' and called it the team's top priority, with active investigation underway
Severe quota exhaustion across tiers
Max 5 plan user ($100/mo) burned through quota in one hour of work; Pro plan subscriber ($200/yr) got only 12 usable days out of 30
Three converging factors
Anthropic reduced peak-hour quotas (affecting ~7% of users), a doubled off-peak usage promotion expired March 28, and suspected caching bugs all hit simultaneously
Caching bugs allegedly inflate costs 10-20x
A developer claims reverse-engineering the Claude Code binary found two independent bugs silently breaking prompt cache; downgrading to v2.1.34 confirmed by multiple users to reduce consumption
Prompt cache lifetime trade-offs
Default cache lifetime is 5 minutes; 1-hour extension doubles write token costs; any brief interruption resets caching benefits entirely
Automated workflows face silent budget drain
Rate-limit errors surface as generic failures and trigger silent retries; a single looping session can exhaust an entire daily budget in minutes
Usage limits remain opaque
Anthropic does not publish exact quotas; Pro plan promises only 'at least 5x free plan usage per session' with no specific numbers, making planning difficult
Detailed breakdown of Claude Code quota exhaustion crisis
What This Means
Developers building automated or agentic workflows on Claude Code face real operational risk from opaque, shifting quota limits and potential caching bugs that silently inflate costs. Until Anthropic resolves the caching issues and clarifies usage limits, practitioners should implement explicit rate-limit error handling, avoid looping sessions, and consider pinning to older client versions. This episode highlights the broader reliability gap between AI vendor marketing promises and the practical constraints of current pricing and infrastructure models.