Effort is the dial that decides how hard a model thinks before it answers you. Turn it up and the model spends more tokens and takes longer to reply. Turn it down and the answer comes back almost straight away.
Effort: the setting that controls how many tokens a model spends reasoning before it replies. Claude ships on high, drops to low, pushes up to max, and the same menu turns thinking off altogether.
Where the effort setting lives
In Claude, click the model name sitting next to the box where you type your message. Scroll to the bottom of that menu and you will find an effort tab. The default is high. You can drop it to low or push it up to max, and the same menu lets you turn thinking on and off.
Effort applies to anything you send: a one-line question, a long document, a task you hand to an agent.
Is effort the same as reasoning?
Effort is how much reasoning you are buying. Reasoning is the model arguing with itself before it commits to an answer, and the reasoning lesson covers that loop in full.
What moves when you turn the dial is how much of the loop runs. On low the model can go from your question straight to the answer, which is what older models did. On max it reasons as far as it is capable of, then checks the answer it came up with, goes back, and keeps working the problem.
Why max is the wrong default
The obvious move is to leave it on max and always get the best result. Two things stop that being a good idea.
It takes longer. And it uses more tokens, which is the part that goes under the radar. The thinking a model does before it replies is made of tokens too, so a max-effort answer costs more than a low-effort one even when the reply on your screen is the same length. The reasoning lesson puts the gap at 200 tokens for a straight answer against 3,000 spent on the reasoning alone: same question, same architecture, a slightly better answer, roughly 15 times the tokens.
Models today are already intelligent enough that pushing a basic task to max works against you. Asking for the warranty length on a boiler does not need max. Low would do it.
| Setting | Use it when |
|---|---|
| Low | Speed matters more than depth, or the question has one obvious answer. Looking up a part number, drafting a two-line reply to a tenant. |
| High | The default, and the right place for most work. Leave it here unless you have a reason not to. |
| Max | Complex tasks where you need high-quality output first time. Working out why a run of solar installs keeps slipping, or planning a week of dispatch across your HVAC vans. |
Does higher effort cost you money?
Inside a chat subscription, no. You pay for it in waiting, and in how quickly you reach the ceiling of your plan. Through an API you pay per token, so effort goes onto the bill directly: subscription vs usage prices a task at $0.60 to $1.00 a run, and the reasoning tokens are the share of that number effort moves.
What turning thinking off does
Off is not a lower rung on the same ladder. It removes the loop, and low can send the model from your question to the answer anyway, so in practice the two land close together: an instant reply, and quality that drops on anything hard. That is the old kind of model, and it is a fine way to run a lookup.
Where the setting earns real attention is work you are not sitting and watching. An orchestrator splitting a job across other agents is doing the multi-step planning that reasoning exists for, and a bad split is expensive: everything downstream executes cleanly against the wrong brief. That is the run where you leave the effort high and pay the tokens.
