The Quota Crisis

Starting around March 22-23, Claude Code users on Max plans — the $100/month and $200/month tiers — began reporting that their usage quotas were draining at dramatically accelerated rates. Users who previously ran 50+ concurrent sessions per day without hitting limits are now exhausting 100% of their quota in approximately 2.5 hours. Simple conversations that should consume a fraction of a percent are burning 2% per message exchange.¹

As of publication, Anthropic has not issued a public statement. The company's status page shows Claude Code as fully operational. GitHub issues filed by affected users are being closed as "duplicates" or marked "invalid" — classified as API-level quota management issues rather than Claude Code bugs.²

The Receipts

The most detailed forensic report comes from a Max 20x plan user ($200/month) who documented the anomaly with token-level precision.¹ The numbers tell a clear story:

What's working normally: Caching shows a 92.7% hit rate. Per-call token breakdowns look correct — roughly 236K cache-read tokens (99.9% of context), 100-500 cache-creation tokens, 1-3 input tokens, and 100-300 output tokens per exchange. The local telemetry says everything is fine.

What's broken: The usage meter climbed from 11% to 18% in twenty minutes of light conversation. At that rate, the full quota exhausts in approximately 2.5 hours. Before March 23, the same user ran multi-agent teams — five or six Opus agents, 50+ sessions per day, 100 million+ tokens daily — without hitting limits.¹

The user's conclusion: per-call token counts appear correct, but the quota decrement rate doesn't match. The most likely explanation is a server-side change in how tokens are weighted against the Max plan budget — possibly cache-read tokens now counted at full rate, or reasoning tokens counted differently.¹

The Pattern

This is not one user's bad luck. At least seven GitHub issues were filed in a 48-hour window reporting the same symptoms: #38350, #38357, #38345, #38335, #38330, #38064, and #37856.² All describe the same pattern — sudden, dramatic acceleration in quota consumption with no change in usage behavior, appearing around the same date.

On Hacker News, users have been reporting related quota anomalies for weeks. "Phantom usage" — waking up to find 20% of a quota consumed despite no overnight activity — was reported as early as March 16.³ Another user reported that the CLI burns 1-3% of quota on startup alone, "even with NO prompts."⁴ A third reported watching their entire weekly limit evaporate during a single Claude Code command execution.⁵

The ecosystem response is telling. Developers have built multiple tools specifically to manage Claude Code's quota constraints: auto-retry scripts that wait for rate limit resets, multi-LLM routers that redirect simple prompts to cheaper models (claiming ~60% cost reduction), and scheduling tools that queue work around limit windows.⁶ When users are building infrastructure to work around your pricing, the pricing is the product's biggest feature.

Bug or Feature?

The community is split. Some users believe this is a bug — an unintended regression in how tokens are weighted against plan budgets. The timing (appearing suddenly around March 22-23) and the lack of any announcement support this interpretation. Bugs happen. Server-side quota calculations are complex. It's plausible.

Others suspect it's intentional — a quiet tightening of how much compute Max plan users actually receive for their $100-$200/month. Anthropic has never published specific token limits for Claude Code plans, describing them only as "5x" or "20x" relative to the Pro plan's unspecified baseline. When your quota is defined as a multiple of an undisclosed number, any change is invisible by design.

A commenter on Hacker News captured the frustration: Anthropic's lack of transparency regarding quota limits "undermines the professional tool reliability needed for business contexts."⁵

The company's response so far — closing issues as duplicates, marking them "invalid," and maintaining a green status page — is not helping either interpretation. If it's a bug, say it's a bug. If it's a policy change, say it's a policy change. Silence invites the worst-case reading.

The Bigger Picture

This is not the first time an AI coding tool has abruptly changed the deal. In the week of March 16-22, 2026 alone: Windsurf killed its credit system without warning, GitHub Copilot locked out paying subscribers during a billing migration, and Cursor shipped a model swap that users didn't request. We covered that chaos in last week's Vibe of the Vibe.

The pattern is consistent: AI coding tools are priced to acquire users, then repriced once users are dependent. The switching costs are high — your prompting style, your muscle memory, your workflow integration, your team's tooling. Every week of adoption is another week of lock-in. And when the quota changes, you don't switch. You complain on Reddit, file a GitHub issue, and keep paying.

We wrote about this dynamic in The Token Squeeze: companies simultaneously pushing developers to use more AI and asking them to spend less on tokens. The squeeze just got tighter.