[BUG] Claude Code Opus 4.6 has become extremely dumb and expensive for serious programming tasks

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Since the February/March 2026 updates (introduction of Adaptive Thinking + default effort changed to medium), Claude Code (especially Opus 4.6) has become significantly worse at real-world programming.
It now:

Makes superficial changes without properly reading the full context or related files.
Ignores instructions from CLAUDE.md and the project's coding style.
Produces code with silly, obvious bugs that didn't happen before.
Declares tasks "complete" when they are only partially done or broken.
Requires constant corrections, creating many extra iterations.

If I leave it on default settings, the output is lazy and low-quality.
If I force /effort max or disable adaptive thinking + set high effort, the quality improves, but it consumes tokens extremely fast — often 2x to 5x more than before.
As a developer who works on complex tasks all day (large refactors, multi-file changes, debugging, architecture), this forces me to choose between bad results or burning through my entire plan quota in just a few days.

What Should Happen?

For complex coding tasks, the model should reason deeply by default (or have a sane default for heavy users), without requiring manual /effort max on every interaction. Quality should not have regressed this much, and the cost/effort trade-off should be reasonable for paying users doing serious work.

Error Messages/Logs

Steps to Reproduce

Open any medium-to-large codebase (mine has ~150+ files with multiple modules, shared utilities, and complex business logic).
Start a new Claude Code session with the default settings (no custom settings.json, adaptive thinking enabled, effort level = medium or high as default).
Give it a realistic complex task, for example:
"Refactor the user authentication flow to support multi-factor authentication while maintaining backward compatibility."
"Debug and fix this failing feature: [describe a real bug you're seeing]."
"Add a new endpoint with proper validation, error handling, logging, and update all related services."

Observe the behavior:
Claude makes edits without reading related files or the full context.
It applies "simplest possible" changes that break existing functionality.
It ignores instructions in CLAUDE.md (style, error handling, etc.).
It marks the task as complete even when there are obvious bugs or missing parts.

To get better quality, run /effort max (or configure CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 + high effort).
Compare token usage: With default settings the quality is poor. With max effort the quality improves but tokens are consumed much faster (often 2-4x more per session), quickly exhausting the weekly quota on Pro/Max plans.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

It was noticeably better before the Adaptive Thinking changes and the default effort adjustment in early 2026 (roughly before February 2026).

Claude Code Version

2.1.109 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_