How hard Claude thinks

There are two dials here and they are constantly confused for each other. One controls how hard Claude thinks before it answers. The other controls how fast the words come back. They’re independent: you can have careful and slow, quick and shallow, or — at a price — careful and reasonably quick.

The thinking dial is the interesting one. Left alone, a model answers with whatever fits immediately, which is fine for most requests and quietly unreliable on anything with several dependent steps. Given room to work through the problem first, the same model gets noticeably more of those right. Same capability, two front doors: a toggle in the Claude app, an /effort level (low, normal, high, xhigh) in Claude Code. On Claude Opus 4.8 the default is already high, so you’re usually reaching for the dial to go up on something thorny rather than to switch thinking on.

Knowing when it pays is the whole skill. Turn it up for problems with a right answer you can’t verify by glancing — a multi-step calculation, a plan with constraints that interact, a decision with real trade-offs, a bug whose cause isn’t where the symptom is. Leave it down for the quick, obvious, one-move requests, where extra deliberation buys you nothing but a wait.

The speed dial, /fast, is a different bargain entirely: it costs a bit more per token and returns answers roughly two and a half times quicker. Crucially it does not make Claude more careless — it changes how fast you get the reply, not how much thought went into it. It’s for the stretches where you’re iterating and the waiting is the bottleneck.

Both dials show up on the bill, in opposite ways: more thinking means more tokens, and fast mode is a straight surcharge for speed. Neither is a substitute for a well-framed request — see prompting — and neither turns a guess into a fact, which is still a trust and accuracy question.