Some language models are built to reason extensively before producing an answer. Prompting them well differs slightly from prompting standard chat models.
Focus on the Goal
Describe the problem, constraints and what a good answer looks like. Detailed step-by-step instructions on how to think are often unnecessary, and can even constrain a model that plans well on its own.
Provide Complete Information
Reasoning models make the most of rich context: requirements, data, edge cases and success criteria.
Specify the Output
Be clear about format and length of the final answer, which is separate from the model's internal reasoning.
Reasoning Budgets
Some APIs let you control how much the model reasons. More reasoning helps hard problems — maths, planning, complex code — but costs more time and tokens. Match the budget to difficulty.
When to Use Them
Complex analysis, multi-step planning, difficult coding and problems where accuracy matters more than speed. For simple extraction or classification, a standard model is usually faster and cheaper.
Verify Outputs
Extended reasoning improves accuracy but doesn't guarantee it. Check important results.
Follow Provider Guidance
Recommended practices differ between model families and change as models evolve; consult the provider's documentation.