Models that think before they answer — and what we use them for in agent architectures.
Category · Agentics
When the model thinks before it answers.
Reasoning models (Claude Opus, GPT-5, the o-series, Gemini 2.5) run an internal "thinking step" before the actual response — structured decomposition, hypothesis forming, self-correction.
For complex multi-step problems — code, mathematics, multi-stage research — that makes a measurable difference. For simple classification, it's rarely worth the overhead.
Where we use it.
In agent architectures for planning steps: what do I need to do, in what order, with which tools. For pure execution — classification, translation, rephrasing — smaller models are enough.
See also
