Claude Code fixes prompt cache misses in 2.1.88 after users report quota spikes

• Claude Code users spent March 30 reporting abrupt quota blowups, including sessions that hit limits after a handful of prompts or a lightweight design chat, which pointed to a product regression more than ordinary heavy usage user report cache bug thread.
• The strongest technical explanation was prompt cache invalidation: users and a GitHub bug report described conversation history dropping out of cache, which can make later turns far more expensive cache bug thread GitHub bug report.
• Claude Code 2.1.88 shipped the next day with fixes for prompt cache misses in long sessions, a --resume crash, and a StructuredOutput schema cache bug that reportedly caused about 50 percent failures in multi-schema workflows 2.1.88 release changelog thread.
• The release also cleaned up misleading quota messaging. One fix changes a false "Rate limit reached" error into the actual entitlement error, which matters because users were clearly struggling to tell policy limits apart from client bugs changelog thread community reaction.
• By early March 31, at least one prominent power user said the rate limit issue seemed fixed, suggesting Anthropic either repaired the caching path quickly or adjusted service-side behavior while the CLI patch rolled out user says fixed 2.1.88 release.

The interesting part is how fast this story snapped into focus. Users first complained about quotas exploding user report, then a reverse engineered theory tied the spike to broken cache reuse cache bug thread, then Anthropic shipped a release with explicit prompt-cache fixes in the changelog via the official release, the changelog entry, and bug threads like conversation history invalidated on subsequent turns. For engineers using long sessions, subagents, and resume flows, the real story here is simple: cache correctness is product reliability.

The most credible explanation for the quota spike was not that Anthropic suddenly made Claude Code dramatically stingier. It was that cached context stopped being reused consistently.

That theory lines up with the public bug report in issue #40524, where a user described conversation history getting invalidated on later turns, leaving only the system prompt cached and forcing large cache rewrites. It also lines up with a separate issue #40851, where a Max $100 user said a mostly text-only session with about 15 prompts consumed 93 percent of quota.

In practice, that is the kind of bug power users notice immediately. If your workflow depends on long-running sessions, background agents, or --resume, cache misses do not just make the tool slower. They make the quota meter look insane.

Anthropic's 2.1.88 release reads like a direct response to exactly that failure mode. The headline fix is explicit: prompt cache misses in long sessions caused by tool schema bytes changing mid-session are fixed in the release notes.

The rest of the changelog matters too:

• Changelog entry: fixed prompt cache misses in long sessions.
• Fixed a StructuredOutput schema cache bug that reportedly caused about 50 percent failures in multi-schema workflows changelog thread.
• Fixed a --resume crash when transcripts contained tool results from an older CLI version or an interrupted write changelog thread.
• Fixed misleading "Rate limit reached" messages when the API actually returned an entitlement error changelog thread.
• Fixed /stats undercounting tokens by excluding subagent and fork usage changelog thread.

That is a bigger deal than it sounds. Broken caching inflates cost. Broken error labels send people chasing the wrong root cause. Broken stats make it harder to prove what happened. 2.1.88 hits all three.

The people who felt this earliest were exactly the users pushing Claude Code hardest. One engineer described Claude Code as nearly unusable under new request-per-minute limits for concurrent agents, and said the missing features in Codex that still kept Claude Code in the mix were session search, looping, and hooks power-user workflow complaints.

That context matters because these are the workflows most sensitive to cache correctness:

• long project histories
• concurrent or swarm-style agents
• periodic check-ins and looping automation
• resumed sessions with lots of prior context
• hook-driven tool orchestration

When those flows break, quota pain compounds fast. A lighter workaround also emerged from the community: switch from Opus to Sonnet, where one user claimed far lower session usage for similar work Sonnet workaround. That is useful tactically, but it is not the fix. The fix is making cached history stable again.

By the morning of March 31, one of the most vocal affected users said the problem seemed resolved, though he was unsure whether Anthropic had fixed a harness bug, changed policy, or adjusted his account specifically user says fixed. That uncertainty is annoying, but the timing is hard to miss.

Users spent March 30 reporting blown quotas user report, the cache invalidation theory gained traction cache bug thread, and the 2.1.88 release landed with cache and schema fixes 2.1.88 release. For engineers deciding whether to trust long Claude Code sessions again, that sequence matters more than the drama did.