Signal, reasoning

The setting most people skip

Reasoning mode ships on by default now, but most people never actually flip it on. Here is what it does and when it earns its keep.

The Curio Team · 3 min readWritten with AI from the sources below and checked against them before publishing.

Reasoning mode has quietly become standard. Claude, ChatGPT, and Gemini all ship a version of it. Yet most people we talk to have never turned it on once. That's the gap we want to close, because the setting does real work.^{1 2}

Here's the plain-English version. When extended thinking is on, the model writes out a private scratchpad before it answers you. It plans, checks its own steps, and catches contradictions it would otherwise sail past. The visible reply comes second. You're paying for a slower first token in exchange for a more careful final one, and on hard questions that trade is worth it.¹

So when does it help? Multi-step problems are the clearest case. Think tax math with several rules stacked together, debugging code where the bug hides three functions away, or a careful contract draft where one clause has to agree with another. Reasoning also shines on anything with arithmetic, logic puzzles, or a chain of dependencies. The model literally shows more of its work to itself, and fewer sloppy jumps survive.^{1 2}

You're paying for a slower first token in exchange for a more careful final one.

And when does it not help? Quick lookups. If you want a capital city, a synonym, a one-line definition, or a fast reformat, reasoning just adds latency and cost for no gain. A short factual pull doesn't get more correct because the model deliberated about it. We leave it off for those. We turn it on the moment a task has more than one moving part.³

The most important thing we can tell you: this is a setting, not a prompt trick. You don't need to type think step by step or beg the model to be careful. In the Claude and OpenAI developer tools it's a parameter you set, and in the consumer apps it's a toggle or a model picker near the message box. Flip it once. The behavior is built in, not summoned by magic words.^{1 2}

Turning it on differs by tool, so it helps to know where to look. In Claude and the OpenAI developer tools it's a parameter you set on the request. In the consumer apps it's usually a toggle near the message box, or a separate model in the picker, sometimes labeled thinking or reasoning. Once you find it, it stays on until you change it, so this is a one-time bit of setup, not a habit you have to remember.^{1 2}

There's a real cost to weigh, and we won't pretend otherwise. Reasoning spends more tokens and takes longer, so it isn't free. For a quick question that trade is a waste. For a decision that matters, a few extra seconds and a few extra cents buy a noticeably more careful answer. We reach for it when being wrong would cost us more than the wait.³

One practical note on budget. Extended thinking uses more tokens, and in most tools you can cap how much thinking is allowed. Give a genuinely hard problem a generous ceiling. Keep everyday chat lean. That single dial, matched to the task in front of you, is most of what good reasoning-mode usage comes down to. Try it this week on your next messy problem and watch the difference.¹