05 of 12
Reasoning
Predicting the next word is fast. Reasoning is when the system takes extra steps: scratch work, checking, trying another path. Some problems are not in the first guess, so the system needs extra time.
Chess and go programs got stronger, with the same trained evaluation, when they thought longer at move time and searched among futures. In 2022, asking a language model to show its working often helped on math and logic, because the extra text became notes the next guess could read. Labs now treat how long a system thinks as something you pay for.
Is a long answer the same as thinking?
In 2022, Jason Wei and colleagues at Google showed that writing out the working, with a few worked examples in the prompt, lifted scores on math and logic. The extra words were notes the next guess could read. If you never run the program or open the cited page, the long answer can still be wrong.
Why does thinking longer cost more?
In 2024 the labs started selling extra thinking time as a product. The machine is still producing pieces of text, some of them hidden. More steps means more computer time after you ask, and more waiting.
When does extra time not help?
François Chollet's little games stayed near zero in an ordinary chat, no matter how long the reply. The jump in 2026 came when the setup kept private notes between turns.