Reasoning and thinking models

Early on, people discovered that asking a model to "think step by step" improved answers to maths and logic problems. Writing out intermediate steps gives the model more text to build on before committing to an answer. This is called chain of thought.

Today most leading models have this built in. They are called reasoning or thinking models: before the visible answer, they generate a stretch of reasoning tokens in which they plan, try approaches and check their work. OpenAI's GPT models, Anthropic's Claude models and Google's Gemini models all offer it, and many open models do too.

What you should know:

• You usually do not see the raw reasoning. Providers return nothing, a short summary, or an encrypted version.
• You still pay for it. Reasoning tokens are billed like output tokens and count towards the token limits, so set the maximum output high enough for thinking plus answer.
• There is a dial. Settings called reasoning effort, thinking level or similar (for example low, medium, high) trade quality for speed and cost. Some models decide by themselves how much to think.
• Temperature is often unavailable. Reasoning models usually ignore or reject a custom temperature setting.

When to use more reasoning: multi-step problems, analysis, planning, coding, tricky instructions, and agent tasks that use tools. When to use little or none: quick classification, rewording, simple extraction, chat replies where speed matters.

Prompting changes slightly. You no longer need "think step by step" for a thinking model. Describe the goal and what a good answer looks like, rather than prescribing every step; high-level guidance usually works better than a rigid script.