Some AI tools now let you choose how hard the model works on your question. There’s usually a toggle labeled something like “thinking,” or a model picker where one option is obviously slower than the others. Picking one is a real decision, and it costs real money.
That setting is reasoning effort. It’s the closest thing to a volume knob any of these products have.
What it means
Reasoning effort controls how much the model thinks before it starts writing you an answer. Those thoughts aren’t free or invisible to the billing system: they’re tokens the model generates, mostly behind the scenes, and you pay for them. More effort buys more of that internal work, which usually helps on hard problems and does nothing at all on easy ones.
Why it matters
Thinking tokens get billed at the output rate, the expensive kind. Turning effort up on a task that never needed it is money spent on nothing. Translating a paragraph or pulling a date out of an email doesn’t get more accurate because the model deliberated about it first.
The dial is a request, not a guarantee. On some models it works as a ceiling rather than a floor, so you can ask for maximum effort and still get a fast, shallow answer because the model decided your question was easy. Every vendor also means something different by “high.” OpenAI ships named effort levels, Google’s Gemini models use thinking levels, and Anthropic spent a while measuring it as a raw token budget before moving to low, medium, and high too. There’s no shared unit for any of it.
Speed turns into a product decision. If you’re putting AI in front of a customer who is sitting there waiting, high effort can stretch a two second reply into most of a minute (ish, depending on the model and the question). Fine for a contract review that runs overnight. Bad for a chat widget.
Simple example
You hand a colleague a forty page lease and ask them to skim it before lunch. You get an answer in twenty minutes and it’s probably fine.
Hand the same person the same lease and tell them to take the whole afternoon, and you get a better answer, four hours later, and their afternoon is gone. Neither version is the right one. It depends on what’s in the lease, and you don’t always know that before you ask.