Claude Code's Hidden Setting That Restores Pre-Nerf Reasoning Quality

Anthropic silently changed Claude Code's default effort level from high to medium back in March. Most users have no idea.

One team measured a 67% reasoning quality drop across 6,852 sessions after the change. If you've noticed Claude Code making more mistakes lately, this might be why.

the fix

Add this to your shell profile (~/.zshrc or ~/.bashrc):

export CLAUDE_CODE_EFFORT_LEVEL=high

Restart your terminal. That's it.

what the effort levels actually do

Claude Code has thinking budget tiers that control how much reasoning the model does before responding:

• low: minimal thinking, fastest responses, cheapest on token usage
• medium: the current default (was high before March)
• high: what most power users expect, the pre-nerf default
• max: maximum reasoning depth, significantly more tokens

For daily coding work, high is the sweet spot. It's what Anthropic originally shipped as the default before quietly dialing it back.

when to use max vs high

max sounds better but it causes problems on routine tasks. The model starts overthinking simple refactors and burns tokens on reasoning that doesn't improve the output.

My rule of thumb:

• high for everything by default
• max only for genuinely hard problems (architecture decisions, complex debugging across multiple files)
• You can switch per-task with /effort max in the CLI

the thinking keywords

You can also nudge thinking depth per-prompt with keywords:

• "think" triggers ~4K thinking tokens
• "think hard" or "megathink" triggers ~10K
• "think harder" or "ultrathink" triggers ~32K

These stack with the effort level setting.

if you're on opus 4.7

4.7 always uses adaptive reasoning regardless of the CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING env var (which only works on 4.6). For 4.7, the effort levels are your only control surface.

The community consensus so far: xhigh is the sweet spot for 4.7 coding work. max works but leads to overthinking.

bottom line

If you haven't touched your effort level settings, you're running on a nerfed default. One line in your shell profile fixes it.

---

Developer working on AI agent infrastructure. Previously: How I Cut My Claude API Bill 60% and Opus 4.7 Token Analysis