Opus model quality regression — significant degradation starting ~April 15, 2026

What happened

Starting approximately April 15, 2026, Claude Opus 4.6 in Claude Code has shown a noticeable quality regression compared to the previous month+. The same model that was working well for weeks is now significantly worse.

Symptoms

• Fails to understand user instructions that were previously handled well
• Ignores CLAUDE.md project instructions — the 4-tier CLAUDE.md hierarchy (global → workspace → functional → project) that was being followed reliably is now partially or fully ignored
• Doesn't update memory/MD files when instructed to — a previously reliable behavior
• Overall reduced comprehension and follow-through — feels like a less capable model than what was running before
• Skill dilution worsened — model was already showing "heavy-chat skill dilution" (documented Apr 7-8), but this is now worse even in fresh sessions

Context — Heavy Power User

I'm a heavy Claude Code power user (Max 20x plan) with:

• 100+ custom skills across multiple workspaces
• 26 active MCP servers (biomedical research + dev tools)
• 4-tier CLAUDE.md hierarchy with detailed project instructions
• 90+ memory files tracking feedback, project state, and user preferences
• Multiple complex workspaces (research papers, web apps, Android games)

This setup has been working well for over a month. The model was reliably reading CLAUDE.md, following skill instructions, updating memory files, and producing high-quality output.

Timeline of Known Issues

• Mar 31 – Apr 7: Token burn bug (prompt-cache attestation, acknowledged by Anthropic as "top priority"). Partially fixed in v2.1.90.
• Apr 7: Documented "heavy-chat skill dilution" — Opus skips output sections in long sessions. Mitigated with lighter skill splits and Completion Gates.
• Apr 15: 3-hour outage (10:53 AM – 1:42 PM ET). Login failures, usage limit glitches, API errors.
• Apr 15-16: Quality regression noticed — model behavior significantly degraded even in fresh, short sessions with minimal context.

What Changed

• v2.1.104 → v2.1.109 updated on Apr 15
• The quality drop correlates with the Apr 15 update and/or the Apr 15 outage
• Previous versions (v2.1.92 through v2.1.104) worked well with the same workspace setup

Environment

• Claude Code v2.1.109 (VS Code extension)
• Model: claude-opus-4-6 (1M context)
• Windows 11 Pro
• Max 20x plan
• Issue started: ~April 15, 2026

Notes

This is not a one-off — it's a sustained drop in quality across multiple sessions over 2 days, compared to consistent good performance over the prior month. The model behaves as if it's a different, less capable version despite showing the same model ID (claude-opus-4-6).

The quality difference is stark enough that a user with 100+ skills and 90+ memory files — who has been using Claude Code daily since February 2026 — can clearly tell something changed.