What is a reasoning model?

A reasoning model is an LLM built to think before it answers, generating a chain of internal reasoning steps first and then using that scratch work to write its final response, instead of blurting out the first thing that comes to mind.

How is a reasoning model different from a regular LLM?

Regular LLMs answer in one pass: you ask, and the model predicts the next word, then the next, straight through to the end. A reasoning model adds a step in between. Before it commits to an answer, it works through the problem in a "thinking" phase, sometimes shown to you, sometimes hidden, where it breaks a math problem into pieces, checks its own logic, or tries an approach and backs out if it is not working.

Only then does it write the answer you actually see. The trade-off is time: that middle step costs extra computation, so a reasoning model is usually slower to respond than a straight-through one — you're trading a little speed for a more careful answer.

What is test-time compute?

This has a name: test-time compute. Instead of just making the model smarter during training, labs let it spend more time, and more computing power, at the moment you ask it something. That extra thinking noticeably helps with math, coding, and multi-step logic problems, where a fast first guess is more likely to be wrong.

Which models are reasoning models?

OpenAI's o-series, o1 and then o3 and o4-mini, put this approach on the map. Every major lab has since shipped its own version:

• OpenAI — the o-series, o1 and then o3 and o4-mini.
• Anthropic — Claude has an extended thinking mode.
• Google — Gemini has a thinking mode with an adjustable thinking budget.
• DeepSeek — R1 does the same thing as an open-weight model you can download and run yourself.

The pattern is the same across all of them — give the model room to think first — even though each lab brands and exposes it a little differently.