Three months ago, Anthropic quietly nerfed Claude Code's thinking depth by defaulting Opus 4.6 to medium effort. Users noticed immediately. Code quality dropped, complex debugging got worse, and the community pushed back hard enough that Anthropic restored the "ultrathink" keyword in version 2.1.68.
Here's what you need to know about the effort system and how to actually use it.
Claude Code has four effort tiers that control how much reasoning the model does before responding:
- Low: 2-5 seconds. Good for quick questions, simple file edits, straightforward tasks
- Medium: 5-15 seconds. The current default for Opus 4.6. Handles most day-to-day coding fine
- High: 15-60+ seconds. Deeper analysis, better for multi-file refactors and tricky bugs
- Ultrathink: The full 31,999-token thinking budget. This is Claude working through a problem the way you'd want a senior engineer to - considering edge cases, evaluating tradeoffs, catching issues before they become bugs
The practical difference between medium and ultrathink is significant. Medium effort tends to give you the first reasonable answer. Ultrathink gives you the considered answer - the one that accounts for how your change affects the rest of the codebase.
To use it, type /effort high in Claude Code, then add the word "ultrathink" anywhere in your prompt. That combination triggers the maximum thinking budget for that specific turn. It's not a permanent setting - each prompt defaults back to your base effort level.
The tradeoff is real: ultrathink is slower and burns through your token allocation faster. Don't use it for renaming a variable or writing a simple function. Save it for architecture decisions, debugging problems that span multiple files, or reviewing complex pull requests.
One tip that pairs well with this: end important sessions by asking Claude to write a "Summary of learnings" paragraph. Claude Code's auto-memory system picks this up and injects it into future sessions, so you don't lose context between conversations.
The default change to medium effort was probably the right call for most users - nobody wants to wait 60 seconds for Claude to think deeply about adding a console.log. But for the moments when you need Claude's best work, ultrathink is the difference between a tool that codes and a tool that engineers.