Why the newest AI models are slower on purpose

AI used to answer the instant you asked. The newest models often do something different: they pause, work through the problem step by step, and then answer. The pause has a name, reasoning, and it is why the latest models are far better at hard problems.

The improvement did not come from building a bigger model. On a hard mathematics exam used as an industry benchmark, AIME 2024, a fast model answering instantly scored 12 percent. A reasoning model from the same era and the same lab scored 74 percent. The one difference was letting it think before answering.

The mechanism is called test-time compute. Instead of spending all its effort during training, the model spends extra effort at the moment you ask, writing a private chain of thought before it replies. The nearest human equivalent: giving someone sixty seconds to answer instead of five. There is a catch, and it matters for anyone paying for these systems: past a point, more thinking just costs time and money without improving the answer.

So when should you let it think? Let it think on hard, multi-step work: analysis, planning, mathematics, code, decisions with several moving parts. Keep it fast for everything else: quick lookups, simple drafts, routine questions. You are trading seconds for quality, and the trade only pays where quality actually matters.

The skill now is not only writing a good prompt. It is knowing when your question deserves a model that stops to think, and when a fast answer was always enough.