Reasoning Models Explained: Why Some AIs "Think" Before Answering

You may see assistants that pause, show a "thinking" step and then answer. These are often called reasoning models.

A standard model produces an answer straight away. A reasoning model spends extra computation working through the problem step by step before responding, such as breaking a task into parts, trying approaches and checking its work.

• Maths and logic puzzles.
• Complex coding and debugging.
• Multi-step planning.
• Analysis with many constraints.
• Scientific and technical questions.
• Tasks where checking beats guessing: reviewing contracts for contradictions, say.

• Slower: thinking takes time.
• More expensive: more computation per answer.
• Overkill for simple jobs: drafting a friendly email doesn't need deep reasoning.
• Not infallible: they can still make mistakes or build convincing but wrong chains of logic.
• Visible reasoning isn't always the true process: treat explanations as helpful, not as a perfect window into the model.

1. State the goal and constraints clearly. You usually don't need to say "think step by step".
2. Provide the relevant data.
3. Specify the output format.
4. Ask for verification: "Check your answer against these constraints."
5. Don't over-instruct. Let the model plan the approach.

Reasoning models spend more effort to get harder problems right. Use them when accuracy on complex tasks matters, and still verify the results.