Effort level

Effort level is a dial for how hard a model thinks before it answers. Turn it up and the model spends more time reasoning through the problem and checking its own work, which helps on hard math and tricky code. Lower it and you get a faster, cheaper answer, fine for simple stuff. Reasoning models expose this as a setting, from low up to high or max, and Claude Code has a /effort command for it.

More thinking isn't free. Higher effort means more tokens and a bigger bill, and past a point it stops helping. On an easy question, cranking it to max mostly makes you wait for nothing. Save the high settings for hard problems, where the extra thinking can flip a wrong answer to a right one.

See it for yourself. Simon Willison benchmarks models by asking them to draw a pelican riding a bicycle as an SVG, then looking at the picture. It works because you need no expertise to judge the result: at low effort the bicycle comes out as loose shapes and the pelican barely reads as a bird, and at maximum effort the same model produces recognisable wheels, a frame, and a bird sitting on it. Run any prompt you care about two or three times at different settings and you will see the same spread on your own work.

The cost spread is the other half. In published runs of that benchmark, one prompt at maximum effort came to about $0.63 on a large model and about $0.02 on a small one. Thirty times the price is worth it on a decision you will live with for a year, and absurd on a first-draft email.

Turn effort up when a problem is hard and getting it right matters more than speed: gnarly debugging, or math you can't afford to fumble. Turn it down for quick lookups and anything where you'd spot a bad answer instantly, since the extra thinking is wasted there.

A rule of thumb we use in trainings: the defaults are right about 80% of the time. Reach for the dial in two cases. Raise it for work that is especially important, like a plan you will act on or an analysis someone will challenge. Lower it for work that is high-volume or context-heavy, where you mostly need the model to follow instructions rather than think hard.

The trap is leaving it high all the time, which quietly runs up cost and latency for no gain on easy work. Match the dial to the task, and if you're not sure, most tools have an auto setting that picks for you.