What do "thinking" and "effort" settings actually change?
More thinking helps on some jobs and wastes time on others.
Most AI apps now have a setting called something like Thinking, Extended thinking, Reasoning or Effort: low / medium / high.
Flip it on and the answer takes longer. Sometimes it's much better. Sometimes it's the same answer, just slower.
So what's going on behind the switch?
When thinking is on, the model first writes out its working, like scrap paper, and only then writes the answer you see.
Reasoning model: a model trained to work through a problem in steps before answering. The working is often hidden or shown as a collapsed "Thinking…" section.
Remember it writes one word at a time (see "AI isn't looking anything up"). Writing out the steps gives it more words to build on before it commits to an answer.
Ethan Mollick's advice for anything high-stakes: use the most advanced model you can, set to "at least the 'High' thinking levels."
For these, the extra thinking just costs time and usage.
And it has limits. In 2025, Apple researchers found that reasoning models gave up completely once puzzles got hard enough. They even used less effort as the puzzles got harder.
The working counts as writing. Hidden or not, every word of it uses your allowance (see "Why your AI usage runs out so fast").
A high-effort answer to a simple question can use many times what a quick answer would.
In spring 2026, Anthropic turned down Claude's default effort level from high to medium. People noticed answers getting worse before they knew why (see "Is my AI getting worse?").
Defaults change. Check yours when something feels off.
It's tempting to read the "Thinking…" section as a window into its mind.
It's useful, but not fully reliable. Anthropic found that reasoning models often didn't mention a hint they'd actually used. They admitted it only 25% to 39% of the time.
The notes are more text it wrote, not a recording of what happened inside (see "Why AI's apologies don't stick").
Turn thinking up for hard, high-stakes work. Planning, analysis, debugging, anything with several steps.
Turn it down for quick jobs. You'll save time and usage.
Skim the thinking, but judge the answer. The working can help you spot a wrong assumption. It isn't proof the answer is right.