Claude Code v2.1.89 fixes the long conversation cache invalidation issue, and the abnormal consumption problem has been addressed.

According to 1M AI News monitoring, Anthropic released Claude Code v2.1.89 (the npm package is already installable, and the GitHub release page has not been updated yet). It fixes multiple caching issues that can lead to abnormal API costs, along with a batch of stability bugs. Previously, Anthropic engineer Lydia Hallie confirmed that the speed at which users reached the usage limits was “far faster than expected.” The community also reported two caching bugs that can raise API costs by 10–20x.

This update fixes caching-related issues including: in long sessions, changes to the tool schema bytes mid-session cause the prompt cache to become invalid; and nested CLAUDE.md files get injected repeatedly dozens of times during long sessions. Both issues cause token consumption in long sessions to far exceed expectations. In addition, it also fixes a StructuredOutput schema caching bug (previously causing about 50% of workflow failures) and a memory leak triggered by large JSON inputs being retained under LRU cache keys.

Other changes worth noting:

1. Using the Edit tool on very large files (over 1GB) no longer triggers an OOM crash
2. When resuming a session with --resume, it no longer crashes on old-version tool outputs
3. When hitting the limit, it no longer shows the misleading “Rate limit reached”; instead it displays the actual error and recommended actions
4. thinking summaries are no longer generated by default; you must manually enable showThinkingSummaries: true in settings
5. Added CLAUDE_CODE_NO_FLICKER=1 environment variable to enable a flicker-free virtual scroll rendering
6. Added a PermissionDenied hook; in auto mode, after refusing a command, the model can be instructed to retry by returning {retry: true}