$200 a month. That's what Anthropic's Max 20x plan costs for Claude Code, and users on that tier have been hitting their weekly usage ceiling in as little as 70 minutes after reset.
What started as scattered complaints in late March has become the biggest controversy in AI coding tools this year. Anthropic has publicly acknowledged the problem, calling it "the top priority for the team," but the situation is messier than a simple rate limit adjustment.
Three Problems Stacked on Top of Each Other
The rate limit crisis isn't one bug. It's at least three separate issues that hit simultaneously.
First, a deliberate throttle. On March 26, Anthropic's Thariq Shihipar confirmed the company intentionally adjusted limits during peak hours: weekdays between 5am-11pm PT, users burn through their 5-hour session limits faster. The weekly ceiling didn't change, but the distribution across the week shifted. Anthropic estimated roughly 7% of users were affected.
Second, a promotional period ended. Around March 28, a promotion that had doubled usage limits outside a 6-hour peak window quietly expired. Users who had calibrated their workflows to the higher limits suddenly felt the floor drop.
Third, and most damaging, a confirmed prompt caching bug. A security researcher reverse-engineered the Claude Code binary and found two independent bugs that silently inflated token consumption by 10-20x:
- A sentinel string collision in Anthropic's custom Bun fork caused cache prefix validation to break on every request, forcing full cache rebuilds. Instead of reusing cached context (which costs $0.30 per million tokens for Sonnet), every interaction was charged at the full input rate.
- Cloud MCP connectors like Ahrefs and Supabase were injecting complete tool definition schemas into every API call regardless of whether the user had activated them. A single connector could add tens of thousands of unnecessary tokens per message.
The math on the cache bug alone is brutal. Users were unknowingly paying $0.04-$0.15 per interaction in wasted tokens, accumulating to 10-20x inflation over a session.
The Fix Is Partial
Anthropic shipped a patch in version 2.1.91 that reportedly addresses the cache bugs. But five related issues remain unfixed: silent tool result truncation, false rate limiting, context stripping, log inflation, and undocumented server-side changes.
Users have found workarounds. Downgrading to version 2.1.34 showed a "very noticeable difference" in token consumption. Running Claude Code via npx instead of the standalone binary bypasses the sentinel bug entirely. Disconnecting unused MCP connectors eliminates the schema injection problem.
The deeper issue is transparency. Anthropic doesn't publish absolute token limits for any subscription tier - only relative multipliers (5x, 20x versus the base Pro plan). Without concrete numbers, it's impossible for developers to budget their usage or verify whether limits are being applied correctly.
The Competition Is Paying Attention
The timing couldn't be worse for Anthropic. Cursor now offers unlimited Auto mode on all paid plans starting at $20/month. Aider, an open-source alternative, reportedly uses 4.2x fewer tokens than Claude Code for equivalent tasks, with typical monthly costs of $60-80 for heavy usage. Cline offers full bring-your-own-key flexibility with no artificial limits.
For Claude Code users weighing their options: update to v2.1.91, disconnect any MCP connectors you're not actively using, and monitor your usage percentage closely after the update. If the drain continues at abnormal rates, the npx workaround or a temporary switch to Aider may be worth the friction.