Claude Code Quality Issues Fixed: Three Bugs Resolved in v2.1.116

TL;DR

• Anthropic identified and fixed three separate bugs that degraded Claude Code quality between March and April
• Issues ranged from wrong default reasoning effort to a caching bug that made Claude forgetful to a system prompt change that hurt coding intelligence
• All fixed as of April 20 (v2.1.116); usage limits reset for all subscribers

What Dropped

Anthropic published a detailed postmortem explaining why Claude Code felt less intelligent for some users over the past month. Three unrelated bugs—each affecting different models and user segments—combined to create the appearance of broad degradation. All have been resolved as of v2.1.116.

What Happened

The first issue: on March 4, Anthropic changed Claude Code's default reasoning effort from high to medium to fix UI freezes in high-effort mode. Users hated it. The tradeoff wasn't worth it—developers wanted smarter responses, not faster ones. Reverted April 7.

The second issue was nastier. A caching optimization shipped March 26 had a bug: instead of clearing old reasoning once when a session went idle for over an hour, it cleared reasoning on every turn for the rest of that session. Claude would lose context of why it made previous edits and tool calls, appearing forgetful and repetitive. This also drained usage limits faster because cache misses increased. Fixed April 10.

The third: Anthropic added a system prompt instruction on April 16 to reduce verbosity in Opus 4.7 (which tends to be chatty). The instruction—keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail—combined with other prompt changes to hurt coding quality. Internal ablations showed a 3% drop in evaluation scores for both Opus 4.6 and 4.7. Anthropic immediately reverted it on April 20.

Developer Impact

If you're using Claude Code, you're already on the fixed version. No action needed. The caching bug fix means your sessions will maintain context properly again, and you'll see more consistent reasoning across multi-turn conversations. Usage limits should stabilize to expected consumption levels.

Anthropic reset usage limits for all subscribers as of April 23 to account for the inflated token usage caused by the caching bug. If you hit limits unexpectedly during the affected period, you're covered.

For teams running Claude Code internally or via the API: verify your reasoning effort settings are back to high (Sonnet/Opus 4.6) or xhigh (Opus 4.7). If you hardcoded medium effort as a workaround, you can revert that now.

Should You Care?

Yes, if you noticed Claude Code feeling less capable over the past month. The fixes restore the intelligence you expected. If you didn't notice degradation, nothing changes for you—you're already running the corrected version.

The postmortem is worth reading if you're curious about how these bugs slipped through. Anthropic is implementing tighter controls on system prompt changes, broader per-model evaluations, and improved code review tooling to catch similar issues earlier. They're also requiring more staff to use the exact public build of Claude Code during testing, not internal variants.

Source: Anthropic