A reasoning model is a large language model trained to think through problems step by step before giving its final answer. It generates internal reasoning, often called “thinking” tokens, to break down the problem, try approaches and check its work. By spending more computation at answer time, reasoning models perform much better on maths, coding, science, logic and planning than standard models of similar size.
This page covers reasoning models as a type of LLM. For the broader concept, see AI reasoning; for the prompting technique, see chain-of-thought.
Key takeaways
- Reasoning models “think before answering” instead of responding immediately.
- They are trained mainly with reinforcement learning on problems with checkable answers.
- More thinking time, called test-time compute, generally improves accuracy on hard problems.
- They are slower and more expensive per question, because thinking tokens cost money.
- OpenAI’s o1 (2024) and DeepSeek-R1 (2025) brought reasoning models into the mainstream.
A reasoning model explained in simple terms
Ask a standard model a tricky maths puzzle and it answers like someone blurting out the first number that comes to mind. A reasoning model is like someone who takes out scratch paper, works through the steps, notices a mistake, corrects it and only then gives an answer. The scratch paper is the thinking; the final answer is what you usually see.
How reasoning models are trained
- Start from a strong pretrained and instruction-tuned LLM.
- Train with reinforcement learning on verifiable tasks: maths problems with known answers, code that must pass tests, logic puzzles. The model generates reasoning and an answer; correct answers are rewarded.
- Learn useful behaviours: over many rounds, models learn to break down problems, backtrack, try alternatives and verify results, often without being shown exactly how.
- Refine with additional supervised data and preference tuning for readability and safety.
DeepSeek’s R1 paper (January 2025) described how large-scale reinforcement learning produced reasoning behaviours, and DeepSeek released the model with open weights, which accelerated research across the field.
Standard LLM vs reasoning model
| Aspect | Standard LLM | Reasoning model |
|---|---|---|
| Response style | Answers immediately | Thinks, then answers |
| Accuracy on hard maths and code | Lower | Much higher |
| Latency | Fast | Slower, sometimes much slower |
| Cost per request | Lower | Higher due to thinking tokens |
| Best for | Chat, writing, simple Q&A | Complex analysis, maths, coding, planning |
Many providers now offer hybrid models that can answer quickly or think longer, sometimes with an adjustable “thinking budget” or effort setting.
A worked example
Question: “A project has tasks A (3 days), B (2 days, needs A), C (4 days, needs A) and D (1 day, needs B and C). What is the minimum project length?”
A reasoning model’s thinking might go: A takes days 1–3. B and C can both start after A. B ends day 5; C ends day 7. D needs both, so it starts after day 7 and ends day 8. Check: is there any shorter path? No, C is on the critical path. Answer: 8 days.
A standard model might rush and add all durations (10 days) or miss that B and C run in parallel.
Test-time compute
Traditionally, AI gains came from bigger models and more training. Reasoning models add a second lever: spending more computation when answering. Allowing a model to think longer, or to try several solutions and pick the best, often improves results on hard problems, with diminishing returns.
When to use a reasoning model
- Multi-step maths, finance or data analysis.
- Complex coding, debugging and architecture questions.
- Legal or policy analysis against multiple rules.
- Planning with many constraints.
- Agentic tasks that need careful step-by-step decisions.
For simple questions, drafting or casual chat, standard models are usually faster and cheaper.
Limitations
- Higher cost and latency.
- Can overthink simple questions.
- Visible reasoning may not faithfully reflect the model’s true process.
- Still makes errors and can hallucinate premises.
- Prompting style differs: over-prescribing steps can hurt performance.