What is Reasoning model?

How it works

A standard chat model starts writing its answer straight away. A reasoning model first generates "thinking" tokens: it breaks the problem down, tries approaches, checks its work and backtracks, and only then writes the answer. This is the chain-of-thought idea built into the model rather than requested in the prompt.

The behaviour comes mainly from reinforcement learning. The model is given problems with checkable answers (maths results, code that must pass tests) and rewarded for reaching correct answers. OpenAI's o1 announcement in 2024 reported that performance improves both with more RL training and with more time spent thinking at test time. DeepSeek's R1 paper showed openly that reasoning behaviour such as self-checking can emerge from large-scale RL.

Most providers now let you control the effort, for example with a reasoning-effort or thinking-budget setting, and many hide or summarise the raw reasoning rather than returning it in full.

Why it matters

• Better on hard, multi-step problems: competition maths, debugging, scientific questions and agent planning.
• Slower and more expensive: thinking tokens are billed and add latency, sometimes many times the length of the final answer.
• Not always better: for simple lookups, rewriting or chat, a fast non-reasoning model is cheaper and just as good.
• Prompting changes: detailed step-by-step instructions matter less; clear goals and constraints matter more.

Many current flagship models are hybrids that can answer quickly or think longer depending on the request or a setting.

Example

Asked "a bat and a ball cost $1.10 in total; the bat costs $1.00 more than the ball; how much is the ball?", a fast model may blurt "10 cents". A reasoning model sets up the equation, checks that 0.05 + 1.05 = 1.10, and answers "5 cents", at the cost of a few hundred extra tokens.