The Rationing: The Cache

On March 23, Claude Code users started flooding GitHub with reports. Quotas that lasted all day were gone in minutes. Pro Max subscribers paying $200/month were hitting limits in 90 minutes. One user reported saying "Morning" and watching 15% of their 5-hour limit disappear. Anthropic's product lead said they were "actively investigating." A community member reverse-engineered the binary in a weekend and found the answer before the company did.¹

The Silent Change

An analysis of 119,866 API calls spanning January 11 to April 11, 2026 tells the story in four phases.²

Phase 1 (Jan 11 – Jan 31): Five-minute cache TTL only. The one-hour tier hadn't been introduced yet.

Phase 2 (Feb 1 – Mar 5): One-hour TTL exclusively. For 33 consecutive days, ephemeral 5-minute tokens were zero. Developers got used to this. Workflows assumed it.

Phase 3 (Mar 6-7): Transition. Five-minute tokens reappeared for the first time in a month.

Phase 4 (Mar 8 – Apr 11): Five-minute TTL dominant. The one-hour window was gone for most requests.

The cost difference is not subtle. Cache write tokens cost 12.5x more than cache read tokens.² With a five-minute TTL, any pause longer than five minutes:a coffee break, a meeting, reading documentation:causes the entire cached context to expire. The next request reprocesses everything from scratch at the write rate instead of the read rate.

In February, with one-hour TTL, the overpayment rate was 1.1%:near zero. In March, after the regression, it was 25.9%.² At Opus pricing, that's $1,582 in unnecessary costs across the dataset.

An independent analysis of 407,000 API turns from a separate researcher on different machines in different countries confirmed the March 6 date independently.²

What Anthropic Said

A Claude Code team member responded that the March 6 change "makes Claude Code cheaper, not more expensive" and that the pre-March behavior "wasn't the intended steady state."²

The comment received 2 thumbs up and dozens of thumbs down.

A user replied: "What I'm getting from your convoluted post is the change is to reduce costs to Anthropic at the users' expense."²

Claude Code lead Boris Cherny responded on X, identifying two causes: tighter limits during peak hours and expensive cache misses from 1M-token context windows.³

The Bugs

While Anthropic investigated, a security researcher ran the 228MB Claude Code binary through Ghidra and radare2, capturing API requests via MITM proxy. He found two independent bugs.¹

Bug 1: The Sentinel Replacement. Native Zig code in the binary searches for a cch=00000 sentinel pattern in request bodies and replaces it with a deterministic hash. If the literal sentinel string appears in the conversation history:from reading the binary, config files, or discussing the cache mechanism itself:the replacement hits the wrong position, breaking the cache prefix. Every subsequent request is processed from scratch without cache.¹

Bug 2: The Resume Regression. Introduced in version 2.1.69, resuming a session relocates system-reminder blocks from messages[0] to messages[N]. Fresh and resumed sessions produce different message structures, yielding different prefix hashes and separate cache chains. A community interceptor patch eliminated the redundant cache creation entirely.¹

The researcher's conclusion, directed at Anthropic: "If a community member can identify the root cause by reverse-engineering your binary in a weekend, why has it taken over a week without a fix:or even a confirmed diagnosis?"⁴

The Perfect Storm

The March crisis wasn't one problem. It was four, stacked:

1. Cache TTL regression (March 6): 5-minute TTL replaced 1-hour, increasing costs 25.9%
2. Two binary bugs: sentinel replacement and resume regression inflating usage up to 20x
3. Peak-hour throttling (announced March 26): weekday 5am-11am PT sessions have tighter limits
4. 2x off-peak promotion ended (March 28): the temporary doubling that masked the TTL change expired

The promotion is the cruelest detail. Anthropic introduced it on March 13, one week after the TTL regression. For two weeks, doubled limits concealed the cost increase. When the promotion ended on March 28, users experienced the full impact of all four issues simultaneously. The March 23 reports came from users who'd already burned through the promotion buffer.

The Scale

This is happening at a company valued at $380 billion after a $30 billion Series G.⁵ Claude Code is Anthropic's fastest-growing product, with demand consistently outpacing capacity.

The rationing isn't poverty. It's growth outpacing infrastructure. Anthropic engineer Thariq Shihipar estimated ~7% of users would hit limits they'd never hit before.³ With millions of users, 7% is a lot of developers watching their quota evaporate.

The same week, Anthropic cut off third-party tools like OpenClaw from using subscription OAuth tokens, forcing them onto API billing. Users who'd been routing frontier AI through $200/month subscriptions now face significantly higher costs at API rates.⁶

The Pattern

GitHub tightened Copilot limits and retired Opus Fast on April 10. Windsurf raised prices. Cursor uses credit drain. Every provider that priced access below cost to capture market share is now correcting.

What makes the cache story different is the mechanism. Rate limits are visible. You hit a wall and you know it. Cache degradation is invisible. Your session looks the same. The model responds the same way. You just run out of quota five times faster, and the only way to understand why is to reverse-engineer a 228MB binary with Ghidra.

The subsidy era isn't ending with an announcement. It's ending with a five-minute TTL.