Claude Settings Guide

The Effort Dial:
which setting to use, and when

By Dan Donovan, @askdandonovan

Claude Opus 5 shipped on 24 July and turned reasoning on by default. One menu next to your send button controls how hard it works on every message you write. Here's where to put it.

5Effort levels
~2 minTo change it
1Menu to find
$0Extra cost
Do this in the next 10 minutes. Open Claude, start a new chat, and look next to the send button. You'll see Effort with five levels and a Thinking toggle. Set effort to Medium, then run your next three normal tasks there before changing anything else. That's the whole first win. Everything below is why, and when to move it back up.
What you'll walk away with
A default that fits
One level you sit on for normal work
When to go up
The two jobs that genuinely need more
The Thinking toggle
What it does and when to switch it off
A rescue plan
What to change first if answers get worse

What actually changed on 24 July

Anthropic shipped Claude Opus 5, and the price was the headline. Worth knowing what that actually means for you: the per-token cost is the same as the model it replaced, and if you're on a plan rather than the API it doesn't touch you at all. Nothing about your bill changed on 24 July. The part that changes your day is underneath it.

The one lever: you're not turning Claude down. You're stopping it doing deep reasoning to reformat a paragraph. That work comes out of your usage whether the task needed it or not.

The five levels, and what each is for

L
Low
Reformatting, tidying, short lookups, quick rewrites. Nothing to think about, so don't pay to think.
M
Medium: start here
Emails, documents, summaries, notes, first drafts. Anthropic's guidance is that low and medium are unusually strong on this model. Most people never need more.
H
High: the default
Analysis, planning, anything where being wrong costs you something. Genuinely better on hard problems.
XH
Extra high
Long multi-step work, coding, agent-style tasks. Anthropic recommends this one specifically for coding and agentic work.
MX
Max: rarely
Hard problems where correctness beats cost and you don't mind waiting. Can show diminishing returns and overthink simple tasks.
The counterintuitive bit. Anthropic's own migration guidance is to start at High and work downwards, checking where quality still holds. Not to start at Max. Max isn't "the good one". It's the setting for problems that actually need it.

The Thinking toggle

It sits right beside the effort menu, and on Opus 5 it's on by default. Switch it off for genuinely simple work and you skip the reasoning step entirely.

One catch worth knowing: you can only switch thinking off at High effort or below. On Extra high or Max the toggle won't let you. That's the product telling you something: if you wanted thinking off, you didn't need that effort level.

Before and after, on one real prompt

Same prompt, two settings. Paste this and try it yourself:

Prompt 
Rewrite this paragraph so it's shorter and less formal. Keep my meaning, don't add anything new.

At Max, thinking on. Claude reasons about tone, audience, register and a few alternative structures, then gives you the rewrite. The rewrite is good.

At Medium, thinking off. You get the rewrite. That's it.

The output you can see is near enough identical, because shortening a paragraph isn't a hard problem. The difference is entirely in what it cost you to get there. That's the whole point of the dial, and it's why the answer isn't "always turn it up."

Where to actually put it

A working starting point, not a rule. Change it per chat. It takes two seconds.

If it doesn't feel right

You'll know it worked when you get through a normal working day without hitting your limit, and you genuinely can't tell the difference in the answers. If you can tell, you've found the task that deserves High. That's useful information, not a failure.
Still hitting limits after this? The dial isn't the only cause. What you load into a chat costs you before Claude has done any thinking at all: whole inboxes, long documents, and bloated memory are the usual suspects.

Everything here was checked against Anthropic's published guidance on 27 July 2026. Interface labels and model behaviour change. If what you see on screen differs from this, trust your screen.