Reasoning Model

There are computer programs that answer questions posed in normal language. Most of them write their answer directly, word for word, without detours. A reasoning model works differently: it first generates a longer chain of intermediate steps for itself, much like doing calculations on a scratch pad. Only afterward does it formulate the actual answer. The user usually only sees the result, sometimes additionally a shortened summary of the intermediate steps. The name comes from the English word “reasoning”, meaning drawing conclusions.

What the scratch pad delivers

For simple questions, a detour is unnecessary. For tasks with multiple steps, it is decisive. A math word problem, a logic puzzle, or a bug in a longer program can hardly be solved in one go. Anyone who starts writing immediately commits early to a path and no longer notices a thinking error.

Reasoning models, on the other hand, can correct themselves within the intermediate text. They try an approach, discard it, and start over. In tests with tasks from math competitions, success rates rose as a result from a few percent to over eighty percent. This was the reason why this class of models received so much attention starting at the end of 2024.

The price for this is time and money. An answer can take a minute instead of a second. Because providers charge per generated text unit, a request quickly becomes ten times more expensive than with a normal model. For the question of France’s capital, this is not worthwhile.

From example to reward signal

Technically, the same basic machinery lies behind a reasoning model as behind an ordinary language model. It always only predicts which text unit is most plausible as the next one. The difference lies in the training, that is, in the learning phase in which the model shapes its behavior from huge amounts of data.

For reasoning models, a training method is used in which the model itself works through tasks with known solutions. If the correct result comes out, the path taken is reinforced. If nonsense comes out, it is weakened. Experts call this reinforcement learning. Importantly: nobody prescribes the thinking steps to the model. It discovers on its own that lengthy checking and recalculating pays off.

A common misconception is that the model is really deliberating like a human. The intermediate steps are themselves just generated text. They help because the model can use them to put more computational steps into a task. This is why one also speaks of more computation at answer time, as opposed to computation during training.

How to recognize them in products

In chat services, reasoning models are usually listed as a separate selection, often with additions such as “Thinking”, “Reasoning”, or an “o” or “R” in the name. Well-known examples are OpenAI’s o-series, DeepSeek R1, and the Thinking variants from Google and Anthropic. Often a notice such as “thinking” appears during the wait, along with a second counter.

In business news, they mainly appear in connection with data centers. Because each request consumes more computing power, the demand for specialized chips increases. The tech stock price slump in January 2025 had exactly to do with this: DeepSeek showed that a strong reasoning model can also be created with considerably less effort.

In everyday use, the distinction is practically worthwhile. For summaries, phrasing help, or translations, a fast standard model is better because it delivers immediately. For physics problems, proofs, spreadsheet analyses, or debugging program code, a reasoning model is the significantly more reliable choice. Wrong answers still occur here too, just less often.