What is a reasoning model?

You ask a chatbot something hard, like a tricky budget or a logic puzzle. Instead of answering right away, it shows the word "Thinking." Sometimes there's a little fold-out note of its steps. A few seconds later, the answer arrives. That pause is the mark of a reasoning model: an artificial intelligence (AI) model trained to work through a problem before it answers.¹²

An ordinary chatbot writes its answer in one pass, one word after another, with no scratch paper. A reasoning model gets scratch paper. Before the final answer, it writes out working of its own. It restates the question, tries an approach, checks a result, and drops a path that doesn't hold up. Then it writes the answer, drawing on those steps. Anthropic's documentation for its Claude models describes thinking in just those terms.¹

Reasoning models are trained with a method called reinforcement learning: try, get a score, adjust, try again. Over and over. The trick is to practice on problems whose answers a computer can check, like math with a known result, or code that either passes its tests or fails them. When the company DeepSeek described its reasoning model in the journal Nature in 2025, the reward came only from whether the final answer was right. Nobody showed the model how to reason. Longer, step-by-step working emerged because it led to more right answers.³

The extra steps help most when a problem has parts that depend on each other. Anthropic's documentation names four kinds of work where thinking improves the results:

• Math, where one slip in an early step spoils the answer.
• Coding, where a fix has to fit everything around it.
• Analysis, like weighing several options against each other.
• Long tasks with many steps, where the plan matters as much as any one step.

A trip with a budget, two time zones, and a ferry schedule is the everyday version of the same kind of problem.¹

For a quick fact, a short rewrite, or a friendly email, the wait buys you very little. The answer isn't hard to find, so there's not much working to do. And the working has a cost. The steps are made of tokens, the small pieces of text a model reads and writes, so they take time. Anthropic's documentation says the tokens spent thinking are billed, even when you never see them. Some newer models decide for themselves how much to think, and may skip it on easy questions.¹

Steps on the screen can make an answer feel checked. But what you see may be a summary. Anthropic says the thinking it shows for its models is a summary of the reasoning, and never the raw working itself. A model can also reason carefully toward a wrong answer. So the result still needs the same check as any other answer. Shown working is a reason to read closely. It isn't a guarantee.¹

Many chatbots now let you pick, or pick for you. A fair rule of thumb: use the quick answer for simple things, and let it think when the problem has parts that lean on each other. Next time you see "Thinking," open the steps and ask one question of them. Does each step follow from the one before?