Particle.news
Download on the App Store

Anthropic Says Claude Code Regressions Came From Harness Bugs, Not Model Changes

All fixes are now in Claude Code v2.1.116.

Overview

  • In a postmortem published Thursday, Anthropic said three product-layer mistakes degraded Claude Code while the API and base model stayed the same.
  • The default reasoning effort was lowered from high to medium on March 4 and restored on April 7, which made replies faster but less capable.
  • A cache change shipped March 26 mistakenly cleared prior reasoning on every turn, and an April 10 fix stopped the forgetful, repetitive behavior.
  • A new brevity system prompt on April 16 hurt coding quality, and Anthropic reverted it on April 20 after ablation tests flagged a measurable drop.
  • Anthropic advises teams to pin model and effort settings, log prompt and session details, and run canary tasks, since wrapper changes can alter behavior even when model weights do not.