Reasoning models

In one paragraph

A reasoning model is a language model trained to work through a problem before it answers. Instead of writing its reply straight away, it first writes out a working: it breaks the question into steps, tries an approach, checks it, and sometimes goes back and tries another. Only then does it give the answer. The working costs time and money, because every step is generated text, but on problems with several dependent steps, such as a calculation, a plan or a piece of analysis, it reaches answers the same model would otherwise get wrong.

Why it matters

For most of their history, language models answered at a fixed cost: one pass through the question, one answer. Reasoning models add a second control. More computation can now be spent on a hard question at the moment it is asked, rather than only by training a larger model in advance. That changes how a system should be designed: which questions deserve the extra thinking, how long each is allowed to take, and how the answer is checked once it arrives.

How it works

• The idea began as a way of asking. Shown a few worked examples with their intermediate steps, large models solved multi-step arithmetic, commonsense and symbolic problems markedly better than when asked for the answer alone.
• Reasoning models build that habit in through training. In one widely studied approach, published in Nature in 2025, the model was rewarded only for reaching answers that could be checked, such as a correct figure or code that passes its tests. Checking its own work and changing strategy emerged during training without anyone writing examples of either.
• At the moment of answering, the model spends extra computation producing its working. Deciding how much thinking each question receives, rather than giving every question the same amount, has been measured as more than four times as efficient. With the computation matched, a smaller model given time to think outperformed a model fourteen times its size, on questions it could already sometimes answer.
• What a large reasoning model learns can be passed on to smaller models, which brings much of the benefit to models small enough to run on hardware you control.

What the evidence says about the limits

• More thinking is not always better. On puzzles whose difficulty could be set precisely, standard models did better on easy problems, reasoning models did better on problems of middling difficulty, and both failed on the hardest, where the models in fact thought less as the problems grew harder.
• The written working is not a faithful record of how the answer was reached. When researchers planted a hint in a question and the model relied on it, the working often mentioned the hint in fewer than one case in five.
• The working is generated text like any other, so a reasoning model is slower and costs more per answer than the same model answering directly.

What it looks like in practice

Take a question that needs several facts combined and a calculation done in the right order, such as whether a change to a contract moves a customer across a pricing threshold. A model answering directly may reply quickly and confidently and get one step wrong. A reasoning model sets out the steps, works through each one and checks the result against the question before replying. The answer takes longer, so it is reserved for questions that warrant it, and the calculation is still repeated in ordinary code rather than trusted because the working reads well.

How this connects to our work

Each kind of request is routed to a model that suits it, so reasoning models take the questions that need them and faster models take the rest. That routing lives in the gateway of private AI infrastructure. Because the written working is not proof, answers that matter are measured against your own evaluation sets, and arithmetic and rule checks run as ordinary code.