Claude Code Defaults to Medium Effort Now. Here's How to Fix It

Something felt off with Claude Code this week. Plans were shallower. Refactors missed edge cases I'd normally catch in review. Debugging suggestions felt like they came from a different model entirely.

Turns out, Anthropic quietly changed the default reasoning effort to medium for Max and Team subscribers. Less thinking, worse results. That's it. That's the problem.

The fix takes 30 seconds. And the best settings depend on what you're paying.

Key Takeaways#

• Opus 4.6 now defaults to medium effort for Max and Team subscribers (changed in v2.1.68)
• Medium is fine for simple edits but kills quality on complex tasks
• Quick fix: claude --model claude-opus-4-6 --effort max
• Permanent fix: edit ~/.claude/settings.json
• Best settings depend on your subscription tier (recommendation table below)

What Changed#

Starting with Claude Code v2.1.68, Opus 4.6 defaults to medium reasoning effort for Max and Team subscribers. The previous default was high.

This happened alongside the return of the ultrathink keyword, which triggered community backlash when it was removed. But here's the thing: the effort downgrade got way less attention. And it affects every single session, not just the ones where you type a magic word.

There's also a known bug (#30726) where setting effortLevel to "max" in your config gets silently downgraded when you interact with the effort selection UI. You think you fixed it. The UI undid your setting. Fun.

Why It Matters#

Claude Code has four effort levels: low, medium, high, max. They control how much extended thinking happens before you get a response. In my testing, lower effort means:

• Less reasoning depth. Claude spends fewer tokens thinking through the problem.
• Shallower analysis. Edge cases, architectural implications, subtle bugs? Missed.
• Faster responses. That's the trade-off Anthropic is optimizing for.

Renaming a variable? Writing a quick test? Formatting a file? Medium effort is fine. You won't notice.

But planning a refactor across multiple files, debugging a race condition, designing an API? The quality drop is real. I went from getting solid architecture suggestions to surface-level plans that missed obvious pitfalls.

How to Fix the Claude Code Effort Level#

Three options. Quick to permanent.

Option 1: CLI Flag (Per Session)#

Launch Claude Code with the effort level baked in:

claude --model claude-opus-4-6 --effort max

One session only. You'll need to remember this every time.

Option 2: The /model Command (Inside a Session)#

Already in a session? Run /model and use the arrow keys to pick your effort level. Four options: low, medium, high, max.

Resets when you start a new session.

Option 3: settings.json (Permanent)#

Edit your global settings file:

// ~/.claude/settings.json
{
  "effortLevel": "max"
}

This applies to all future sessions. But watch out for bug #30726. If you use the /model UI to change effort during a session, it can silently overwrite your config back to a lower value.

You can also set it per-project in .claude/settings.json at your repo root. Useful if different projects need different effort levels.

Best Claude Code Effort Settings Per Subscription Tier#

Not every tier can run max effort all day. After a few weeks of daily use, here's what actually works:

Here's the thing: effort level burns tokens. Max effort on Opus 4.6 uses significantly more tokens per response than medium. On a $20 Pro plan, that means hitting your limit by lunch. On Max 20x, you won't notice.

If you're on Max 5x or Team: run medium by default, toggle to max before complex tasks with /model. Takes two seconds. Saves you from burning tokens on a file rename.

Make It Stick: Shell Alias#

Want max effort every time without remembering flags? Shell alias:

# Add to ~/.zshrc or ~/.bashrc
alias claude-max="claude --model claude-opus-4-6 --effort max"

Now claude-max launches with the right settings. I keep two aliases:

alias claude-fast="claude --model claude-sonnet-4-6 --effort medium"
alias claude-max="claude --model claude-opus-4-6 --effort max"

claude-fast for quick edits. claude-max when you need Claude to actually think.

If you want the full picture on Claude Code configuration (CLAUDE.md, MCPs, skills, and hooks), I wrote a setup guide that covers everything. And if you're still picking which model works best for your workflow, start there.