How Small Backend Tweaks Broke Claude Code for a Month—And What It Means for AI Reliability

For weeks this spring, developers using Anthropic’s Claude Code watched their trusted coding partner lose its edge. Tasks stalled. Responses turned shallow. Hallucinations crept back. By early April, what started as forum gripes had become a full-blown trust crisis.

Anthropic finally confirmed the cause on April 23: three separate engineering changes—none involving model weight updates—had combined to degrade performance. A lowered default reasoning effort on March 4 cut depth to fix interface lag. A caching bug introduced around March 26 erased short-term memory in long sessions. And a system prompt tweak tied to a new Opus release shaved 3% off coding benchmarks. All three were fixed by April 20 in version 2.1.116, with usage limits reset for affected subscribers.

But the damage was done. Users had logged thousands of sessions showing thinking depth drop by roughly 67% in some workflows. Stella Laurenzo, an AMD sales director, analyzed session logs and confirmed a shift toward faster, less thorough reasoning. Anthropic initially brushed off complaints, suggesting users refine prompts or adjust expectations. The company only moved after a GitHub issue packed with ablation studies and before-and-after comparisons forced a postmortem.

The episode reveals a hard truth for AI infrastructure: small backend tweaks—made for speed or efficiency—can cascade into major user experience failures when not tested against real workloads. Anthropic pledged better monitoring, longer ablation tests, and clearer communication. But for many developers, the memory of that month of quiet decline lingers. They now test every new deployment against older baselines. Some are exploring alternatives.

This wasn’t intentional nerfing. Model weights never changed. The issues lived in the harness—defaults, memory management, instructions. Yet the gap between first reports and full disclosure tested customer loyalty. For an industry selling reliability at scale, that gap matters more than any single fix.