Reasoning model
In shortA reasoning model spends extra tokens on hidden step-by-step thinking before it writes its answer, which helps on hard problems but costs more.
In plain words
A standard model starts writing its answer right away. A reasoning model first produces a stretch of internal working, then the answer. The idea is older than the products: a 2022 paper by Wei and colleagues showed that asking a model to write out intermediate steps improved results on multi-step problems. Reasoning models are trained to do this by themselves.
The providers describe the extra tokens in similar ways:
- OpenAI calls them reasoning tokens. They are not shown through the API, but the usage data tells you how many were used. They are billed as output tokens and take space in the context window.
- Anthropic's extended thinking works in the same way: thinking tokens are billed as output tokens. The thinking you see is a summary. Newer models use adaptive thinking, where the model decides whether and how much to think, guided by an effort setting.
- Google's Gemini reports thinking tokens in its usage data too.
Why it matters
- Quality on hard tasks: Multi-step maths, logic puzzles, planning, tricky code and careful analysis tend to improve.
- Cost and time: Hidden tokens add to the bill and delay the first visible word. See price per million tokens.
- Budget risk: OpenAI warns that if the output limit runs out during reasoning, the response can end with no visible answer. Leave room.
- Not always needed: Simple questions, rewrites and lookups do not gain much. A small standard model is cheaper and faster for those.
- Still fallible: Thinking does not remove errors. A model can reason its way to a wrong answer (hallucination).
Example
You ask: "Offer A costs 1,200 a year with a 10 percent discount for two years. Offer B costs 1,050 a year, flat. Which is cheaper over three years?" A reasoning model works through the years step by step and then gives a short answer. A plain model may jump to a guess. For "Fix the typo in this sentence", the plain model is the better choice.
Common misconceptions
- "I can read its full thought process." Often you get a summary, or nothing. Providers differ.
- "Reasoning means it understands." It means more computation before the answer, not human thought.
- "Always pick the reasoning model." Use it where the problem needs steps. See choose a model.
Questions and answers
What is a reasoning model?
A model that produces hidden step-by-step thinking before its final answer. It tends to do better on hard problems and uses more tokens and time.
Are thinking tokens billed?
Yes. OpenAI and Anthropic both bill them as output tokens, even when you cannot see them in full.
When is a reasoning model worth it?
For multi-step maths, logic, planning and hard code. For simple rewrites or lookups, a standard model is cheaper and faster.
Read next
- Token
A token is a small chunk of text, often a word piece, that an AI model reads, writes and bills.
- Price per million tokens
Price per million tokens is the usual way AI APIs state their rates: what you pay for each million tokens sent in or generated.
- Context window
The context window is the maximum amount of text, counted in tokens, that a model can read and write in one request.
- Large language model (LLM)
A large language model (LLM) is an AI system trained on huge amounts of text to predict the next token, which lets it write, summarise and answer.
- How to choose an AI model
Match the model to the job: small for volume, mid for daily work, top or reasoning for hard problems and local for private data, with a decision tree and a test.