What are reasoning models and when should we use them?
Models that "think" before answering are much more accurate in math, programming, and analysis, but slower and more expensive. Understand when they are worth it and how to make the best use of them.
From "Quick Answer" to "Thinking"
Ordinary language models generate responses word by word and instantly; much like someone answering without a pause. This method is great for conversation and writing, but in problems that require multiple steps of reasoning, small early mistakes lead to a wrong final answer.
Reasoning models have a "thinking" stage before the final response: they break down the problem, test different solutions, check and correct their work if necessary, and then provide the answer.
Famous Examples
- The o series and thinking modes in OpenAI's GPT models
- DeepSeek-R1, the first powerful reasoning model with open weights
- Thinking modes in the Claude and Gemini families
- Mistral's Magistral models and Qwen with thinking mode
Today, many major models are "hybrid": they determine for themselves whether a question is simple and should be answered quickly, or complex and requires more thought.
When to Use a Reasoning Model?
| Suitable | Not Necessary |
|---|---|
| Mathematical and statistical problems | Writing an email or caption |
| Debugging complex code | Translating simple text |
| Analyzing a contract or document with multiple conditions | Summarizing an article |
| Multi-step planning | Simple trivia questions |
| Logical puzzles and decision-making with many criteria | Everyday conversation |
The Cost of Thinking
- Time: The response may take anywhere from a few seconds to a few minutes.
- Cost: In the API, "thinking" also consumes tokens and incurs a cost.
- Overcomplicating simple tasks: For simple questions, a reasoning model sometimes unnecessarily complicates things.
Prompt Engineering Tips for Reasoning Models
- State the goal and success criteria precisely, not the method of solution. These models find the path themselves, and overly step-by-step instructions sometimes tie their hands.
- Provide all necessary information at once; constraints, data, and exceptions.
- Ask it to explicitly state its assumptions, so you can review them.
- For final verification, ask the model to test the answer using a different method.
Conclusion
Reasoning models are not a replacement for ordinary models; they are complementary to them. A fast model is sufficient for writing and conversation; when a problem is truly hard or a mistake is costly, allow the model to "think". To get familiar with related terms, see Chain of Thought and Reasoning Model in the glossary.