Reasoning Model

Scratch paper

If you ask someone 7x8, they answer instantly from memory. If you ask them 47x83, they want scratch paper. Reasoning models, also called "thinking" models, are LLMs that learned to use scratch paper. Before answering, they privately write out steps, check their own work, and only then reply.

That thinking phase is text the model generates. You usually just see a "thinking..." pause, and sometimes a summary of it.

When it matters, and when it doesn't

For math, tricky code, and multi-step planning, reasoning is a huge upgrade. The model catches its own mistakes mid-way instead of committing to a bad first guess.

For "write a birthday message," it's a waste. You pay for thinking time in both money and seconds. That's why fast models exist, and why the newest models decide for themselves how hard to think based on the question.

Almost every current flagship is a reasoning model, marked with the brain badge on the comparison table. The main exceptions are speed-focused and older open models.