Anthropic confirms three system-level bugs caused Claude's reported quality decline in 2026

Developers across GitHub, Reddit, and X reported throughout 2026 that Claude was producing shallower reasoning, stopping tasks early, and making edits without first reading files. A detailed analysis by AMD's Stella Laurenzo, covering nearly 7,000 Claude Code sessions, documented a sharp drop in visible reasoning depth and a rise in unverified edits between January and March. Anthropic published a technical postmortem on April 23 identifying three distinct causes: a default reasoning-effort downgrade on March 4, a context-management bug introduced March 26 that repeatedly cleared session memory, and an overly strict brevity instruction added April 16 that reduced coding quality by around 3 percent. All three changes were rolled back within weeks, with Anthropic clarifying that its underlying models and API infrastructure were not affected. The episode highlights how the full Claude experience depends not just on the core model but also on system prompts, effort settings, and product-layer configurations that can silently degrade output quality.