Reasoning

Reasoning roughly translates to “logical deduction” or “drawing conclusions.” In AI programs that write text, the term means something specific: the program doesn’t output its answer right away, but instead works through the task beforehand in intermediate steps. It breaks a question down into sub-questions and answers them one after another. The difference is the same as between a math problem you guess in your head and one you work out on paper. Programs that operate this way are called reasoning models. They have been the most important development in this field since around 2024.

Why guessing isn’t enough for math

At its core, a language model is a machine that predicts the next word. For language, this works remarkably well, because many phrasings are plausible. But for a calculation, there is only one correct result. Anyone who simply picks the most probable next character is often slightly off — and slightly off is completely wrong when it comes to numbers.

This is exactly where earlier chatbots repeatedly failed. They could talk fluently about physics but couldn’t work through a multi-step task. With reasoning, this changes measurably. In upper-secondary-level math competitions, success rates rise from a few percent to over 80 percent when a model is allowed to write out intermediate steps.

Economically, this is why reasoning comes up in almost every piece of AI news. Tasks with a verifiable solution are valuable to businesses: testing software, checking accounting data, reviewing contracts for contradictions. A model that sounds plausible but calculates incorrectly is useless there. A model that calculates slowly but correctly is worth money.

The thinking step before the answer

Technically, reasoning is based on the model generating a longer text for itself before the actual answer. This intermediate text is called a chain of thought. In it, the model formulates assumptions, works through partial steps, and discards approaches again. Only afterward does it write the answer that the user sees. Many providers hide this chain of thought and show only a summary.

The fact that a model displays this behavior at all is trained into it. It is made to solve thousands of tasks whose solutions are known. If a solution path leads to the correct result, it is reinforced; if it leads to the wrong one, it is weakened. This procedure is called reinforcement learning, in other words learning through reward. After many rounds, the model develops strategies on its own, such as double-checking or starting over.

The price for this is computing time. A reasoning model can need several minutes and thousands of words of internal deliberation for a difficult task. Because providers bill per processed text unit, such an answer is often ten times more expensive than a normal one. That’s why providers offer controls that let you limit the thinking duration. So you consciously choose between fast and thorough.

Reasoning in chatbots and headlines

In everyday life, you encounter reasoning in the well-known chat services. There is usually a choice between a fast model and one that briefly pauses at the notice “thinking.” The latter is a reasoning model. For a question about a cooking recipe, it’s not worth it; for a physics problem or a programming bug, it definitely is.

In business news, reasoning models mainly come up in two contexts. First, in the competition among the big providers, who measure their models against exam tasks. Second, in the electricity demand of data centers, since longer thinking means more computing power per request. Chipmakers' stocks regularly react to such news.

A common misconception persists: reasoning is not thinking in the human sense and is no guarantee of truth. The chain of thought can look clean and still contain a fabricated intermediate step. Reasoning significantly lowers the error rate, but it does not eliminate it. For questions without a verifiable solution, such as historical assessments, the benefit is also much smaller than for mathematics.