What Is a Reasoning Model?
Also called thinking model
Short Answer
A reasoning model is a large language model trained to work through a problem step by step before giving its final answer. It spends extra computation at answer time, called test-time compute, which improves results on math, code and multi-step questions, at the cost of slower and more expensive answers.
- Taught in lesson 82
Why It Matters
Reasoning models handle harder problems, such as complex analysis, maths and multi-step code changes, more reliably than standard models. Choosing when their extra cost and time are worth it is a practical decision AI engineers and product teams make regularly.
How It Works
OpenAI introduced the first widely used reasoning model, o1, in September 2024; DeepSeek-R1 followed in January 2025 with openly published weights. Most major model families now offer a reasoning or thinking mode.
Reasoning models are trained with reinforcement learning to produce useful chains of thought. For simple tasks, a regular model is often faster and cheaper; choosing the right model for each task is part of controlling cost.
A reasoning model generates intermediate thinking before its final answer, sometimes hidden from the user. Spending more computation on this thinking usually improves results on hard tasks, which is why it is called test-time compute. For simple tasks, such as rewriting a sentence or extracting a field, the extra steps add cost and delay without much gain, so many applications route easy requests to a standard model and hard ones to a reasoning model.
Reasoning Model in Practice
Everyday Example
Asked to plan a week of meetings across five time zones with constraints, a reasoning model checks options internally before proposing a schedule.
Industry Example: Pharma
A reasoning model can work through a dose calculation or a statistical analysis plan step by step, which a scientist then verifies line by line.
Where You Meet It
- Model options labelled "thinking" or "reasoning".
- Coding agents and complex analysis tasks.
- Benchmarks on maths and science problems.
- Choices between faster and more careful modes in AI tools.
Key Ideas Around It
- Test-time compute
- Extra computation spent while answering.
- Chain of thought
- Working through intermediate steps.
- Model routing
- Sending each request to the most suitable model.
- Latency
- How long an answer takes; reasoning increases it.
A Common Mix-Up
Reasoning models are not always right. They make fewer mistakes on multi-step problems, but they can still reason confidently to a wrong answer, so evaluation and checks remain necessary.
Check YourselfShould you use a reasoning model to fix spelling in short texts?
Usually not: the task is simple, so a standard model is faster and cheaper with similar quality.
Learn It on ai2hired
S7 · #82How LLMs Are Made
Coming in Season 7, Deep Learning and Transformers. Every ai2hired lesson is free on YouTube.
Questions People Ask
Should I always use a reasoning model?
No. Use one for hard multi-step problems. For extraction, short answers or high volumes, a standard model is faster and cheaper.
Is reasoning the same as chain-of-thought prompting?
Related. Chain-of-thought prompting asks any model to show steps; a reasoning model was trained to do that on its own.
Related Terms
Learn AI From Zero to Hired
Free lessons on YouTube, each one built from what real AI job ads ask for.