AI Explainedby Milad Fahmy

AI Explained

Reasoning models: AI that thinks before it answers

Some AI models stop and think before answering. How reasoning models work, what thinking tokens cost, how to set the effort, and when to use them.

Milad FahmyOct 29, 20262 min read

On this page
  1. How they work
  2. Thinking costs tokens
  3. You can set the effort
  4. When to use them

Some AI models answer instantly. Others pause, work through the problem, and only then reply. Those are reasoning models, and knowing when to use them saves both time and money.

How they work

Before writing the final answer, a reasoning model produces a stream of internal reasoning: it breaks the problem into steps, tries an approach, checks it, and sometimes backtracks. Only after that does it write the reply you see.

OpenAI calls these reasoning tokens. Anthropic calls the feature extended thinking: "Claude thinks against that budget before it starts its final answer." Depending on the provider, you may see a summary of that thinking, or none of it.

Thinking costs tokens

That reasoning isn't free. It's made of tokens, and:

  • you're billed for thinking tokens, as output tokens,
  • they take up space in the context window,
  • and they add time before the answer arrives.

A hard problem might use thousands of thinking tokens for a short final answer.

You can set the effort

Most APIs let you choose how hard the model thinks:

  • OpenAI's reasoning.effort: "Lower effort favors speed and lower token usage, while at higher effort the model thinks more completely."
  • Anthropic's thinking budget: higher budgets allow more thorough reasoning, at the cost of more latency.

Start low and raise it only when the answers aren't good enough.

When to use them

Good fit Poor fit
Tricky coding and debugging Quick factual questions
Multi-step planning Rewriting or formatting text
Math and logic Simple classification
Agent workflows with many steps Chat where speed matters most

A good pattern in apps: route simple requests to a fast model and send only the hard ones to a reasoning model, or turn the effort up only for those.

Fast answer or deep thought: pick the right one.

Sources

#reasoning models#extended thinking#llm#tokens#ai models

Watch