Lesson 5 of 14 · Settings · 3 minutes
Thinking effort: thinking longer vs answering fast
Some ovens can think before they answer. These are called reasoning models. Before writing the answer you see, they write hidden notes to themselves: they try ideas, check them and fix mistakes, a bit like a baker who reads the whole recipe before touching the flour.
Thinking costs tokens
Those hidden notes are made of reasoning tokens. You never see them, but they count as output tokens, so you pay for them and wait for them. Under each answer GrannyBakes shows how much the oven thought.
A real example
Here is the same puzzle on the same oven, once with no thinking and once with high effort:
The message A baker has three jars. One holds only sugar, one holds only salt, and one holds a mix of both. They are labelled "Sugar", "Salt" and "Mixed", but every label is wrong. She may taste one spoonful from one jar only. Which jar should she taste from, and how does she then fix all three labels? Answer in a few short sentences.
Same oven, same puzzle. Only the thinking effort changed.
No thinking
Answers straight awayThinking effort: no thinkingTaste from the jar labelled “Mixed.” Since every label is wrong, it contains only sugar or only salt. If you taste sugar, relabel it “Sugar”; the jar labelled “Salt” must then be mixed, and the jar labelled “Sugar” must be salt. If you taste salt, swap “Sugar” and “Salt” in that reasoning.
2.3 s · 84 in / 76 out tokens · $0.000046
High effort
A long think before answeringThinking effort: highTaste from the jar labelled “Mixed”; since every label is wrong, it must contain only sugar or only salt. If it tastes like sugar, that jar is Sugar; the jar labelled “Salt” is Mixed, and the one labelled “Sugar” is Salt. If it tastes like salt, reverse Sugar and Salt in that arrangement.
2.9 s · 84 in / 163 out tokens · 90 of them thinking · $0.00009
Both got it right. The long think took longer and cost about twice as much, because the puzzle was easy. On a hard problem (a tricky plan, a math question, a long set of rules) the extra thinking is often what makes the answer right.
Picking the effort
The thinking effort setting tells a thinking oven how long it may think. The names depend on the oven, usually none, low, medium, high, and sometimes more:
- None or low for simple chats, short emails and quick sorting jobs.
- Medium is the usual default and a good start.
- High when the oven gets things wrong at medium: puzzles, plans, careful checks.
Ovens that can't think simply leave this setting out.
Granny's rule of thumb: start low, and only turn the thinking up when the answers need it. Taste before and after.