Reasoning models (o1/o3, R1, and similar) are LLMs trained to think before answering — producing a long internal chain of reasoning, then a final response. They trade extra inference time for much better answers on hard problems.
Test-time compute
The insight: instead of only scaling training, you can scale thinking at inference. Give the model room to work step by step — and to check, branch, and backtrack — and accuracy on math, code, and logic jumps.
normal LLM: question → answer (fast, shallow)
reasoning model: question → [long think...] → answer (slow, deep)It's chain-of-thought baked into the model via RL training that rewards reasoning that reaches correct answers — so you don't have to prompt for it.
Trade-offs
| Gain | Cost |
|---|---|
| Much better hard-problem accuracy | Slower, more tokens, pricier |
| Self-correction, planning | Overkill for simple tasks |
Use a reasoning model for genuinely hard problems; a normal model is faster and cheaper for everyday turns. This is the current frontier of how far "just predict the next token" can go.
Related: Prompt Engineering Patterns · Reinforcement Learning from Human Feedback (RLHF) · How Transformers Work