Reasoning models spend extra computation thinking before they reply, working through steps, checking and revising. This makes them stronger at maths, coding, logic and planning, but slower and more expensive per answer. Many let you control how much thinking time to allow, and some show a summary of their reasoning.
How it works
A standard language model starts writing its answer straight away, token by token. A reasoning model first produces a hidden or visible stream of reasoning tokens: breaking the problem into parts, trying approaches, spotting mistakes and backtracking, then writes the final answer. This builds on chain-of-thought prompting, where people found that asking a model to 'think step by step' improved accuracy. Reasoning models go further by being trained for it, typically with reinforcement learning on problems that have checkable answers, such as maths questions or code that must pass tests. The model is rewarded for reaching correct answers, so it learns reasoning habits that work. The key idea is test-time compute: instead of only making the model bigger, you let it spend more computation on each question. Longer thinking generally improves results on hard problems, up to a point. The trade-off is cost and speed, because reasoning tokens are usually billed like output tokens and can run to thousands per answer. Many providers offer a setting for reasoning effort or a thinking budget, letting you choose between quick replies and deeper analysis.
Why it matters
For tasks with many steps, a reasoning model can be the difference between a plausible answer and a correct one. It suits debugging code, analysing data, checking contracts, planning projects or solving tricky maths. For quick tasks like rewriting a caption or drafting an email, a standard model is faster and cheaper. Knowing when to switch saves money and waiting time while getting better results where accuracy matters.
Where you will see it
- Debugging a stubborn piece of code that fails only in certain cases.
- Working through a multi-step maths, physics or finance problem with checks.
- Planning a project schedule with dependencies, deadlines and resource limits.
- Reviewing a long contract for clauses that conflict with each other.
Limits and common misconceptions
- Reasoning models can still be wrong, and a long, confident chain of reasoning can make errors look more convincing.
- They are slower and cost more per answer, which is wasteful for simple tasks like rewriting or short replies.
- Displayed reasoning may be a summary and is not a guaranteed, faithful record of how the model reached its answer.
How it differs from related terms
- Reasoning model vs Large language model (LLM) Every reasoning model is an LLM, but a standard LLM answers immediately, while a reasoning model is trained to think first.
- Reasoning model vs Inference Inference is running any model to get an output; reasoning models deliberately use more inference compute per question to improve accuracy.
Tools that use it
Frequently asked questions
When should I use a reasoning model?
Use one for problems that need several steps and where correctness matters: coding, maths, data analysis, logic puzzles, planning or careful document review. For everyday writing, brainstorming, summarising or quick questions, a standard model is usually just as good, faster and cheaper.
Why are reasoning models slower?
They generate many reasoning tokens before writing the answer you see, sometimes thousands. Each token takes computation, so the response takes longer and, on paid APIs, usually costs more. Many tools let you lower the reasoning effort when you want speed rather than depth.
Can I see what a reasoning model is thinking?
It depends on the provider. Some show the full reasoning, some show only a summary and some hide it. Even when visible, research suggests the written reasoning does not always faithfully reflect how the model reached its answer, so treat it as a helpful explanation, not proof.
Related terms
- Prompt engineering The practice of designing, testing and refining prompts and their surrounding context so an AI model gives accurate, consistent and useful output.
- RAG (retrieval-augmented generation) A method where an AI system first retrieves relevant information from a chosen source, then uses it to generate an answer grounded in that material.
- RLHF Reinforcement learning from human feedback: a training method where people rank AI responses, and the model is tuned to produce the preferred kind.
- Speech to text AI that converts spoken audio into written text, also called automatic speech recognition (ASR) or transcription.
- Synthetic data Data created by algorithms or AI models, rather than collected from real events, used to train, test or fine-tune AI systems.