Reasoning models: AI that thinks before it answers
A reasoning model is a large language model trained to work through an internal line of reasoning, sometimes a long one, before giving its answer. It handles hard problems better, at the cost of slower replies and higher prices.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|---|
| Origin | 2024 |
| Introduced by | OpenAI (o1) |
The plain definition
A standard large language model produces its answer word by word, with no scratchpad. A reasoning model starts with a thinking phase instead: it breaks the problem down, tries an approach, checks it, sometimes backtracks, and only then writes its reply. Think of it as chain of thought built into the model through training, rather than something you have to ask for in the prompt.
A concrete example
You ask: “Here are five people’s calendars. Find a one-hour slot next week when everyone is free, avoiding Mondays.” A standard model may suggest a slot that looks plausible but breaks one of the constraints. A reasoning model works through each constraint, rules out impossible slots and then double-checks its solution. It takes longer, but the answer is more likely to be right.
Key milestones
OpenAI introduced o1 in September 2024 in the post “Learning to reason with LLMs”, explaining that the model is trained with reinforcement learning to use its chain of thought productively. In January 2025, DeepSeek released DeepSeek-R1 along with a research paper showing that such reasoning skills can be drawn out largely through reinforcement learning with verifiable rewards (see RLVR). Since then, most model makers have shipped a “thinking” version of their models.
Why it matters
These models shift part of the effort from training time to answer time, which is the idea behind test-time compute. In practice you trade speed and money for reliability. For an agent that has to plan several actions, that trade is often worth it; for a customer-service chatbot expected to reply within a second, much less so, because of latency.
Common myths
- “It no longer makes mistakes.” It makes fewer on certain problems, but it can still get things wrong, sometimes very confidently.
- “The longer it thinks, the better.” Beyond a certain point extra thinking stops improving the answer and simply adds cost.
- “You need very detailed prompts.” Providers tend to recommend clear, direct instructions instead: there is no need to spell out every step.
To compare the capabilities and prices of reasoning models against standard ones, head to the model comparison tool.
Related terms
- Chain of thought · Getting a model to write out intermediate reasoning steps before its final answer, which helps on multi-step problems.
- Test-time compute · The computing power a model spends while answering, for instance to think for longer or to compare several candidate solutions.
- RLVR (reinforcement learning with verifiable rewards) · Reinforcement learning in which the reward comes from an automatic check, such as a correct answer or passing code tests, not from a judgement.
- Large language model (LLM) · A program trained on huge amounts of text to predict what comes next in a text; it is the engine behind AI assistants.
- Latency · The waiting time between sending a request to an AI model and receiving its answer, from the first word through to the last.
- Price per million tokens · How AI APIs bill usage: one rate per million tokens the model reads as input, and another, usually higher, per million it generates.
Try it
Full glossary
Frequently asked questions
How can I tell if I'm using a reasoning model?
Assistants usually flag it with an option such as “Thinking”, “Reasoning” or “Deep think”, and show a pause or a summary of the thinking before the answer. The model name is another clue: check the assistant's profile page.
Is a reasoning model always better?
No. For simple tasks such as rewording, translating or answering a factual question it adds little and costs more. It shines when the task needs several steps of thought.
Can I read all of its reasoning?
It depends on the provider. Some show a summary, some the full reasoning, some nothing at all. Either way, the text helps you follow the answer without guaranteeing it is correct.
Why does it cost more through the API?
Thinking tokens are generally billed as output tokens, even when they are not all displayed. The same question can therefore use far more tokens than with a standard model.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.