Foundations & Models
Reasoning Model
Instead of answering right away, the model first produces an internal chain of thought: it breaks the problem down, checks intermediate steps, discards dead ends. This costs time and compute at answer time, but markedly raises quality on mathematics, code and multi-step logic. What is new is the trade: you buy care with waiting time — decidable per question.
An example
Asked for the capital of France, a reasoning model answers just as fast as any other. Asked a question that weighs three conditions against each other — which rate is cheapest at this consumption, this term and this notice period? — it visibly works longer, checks intermediate steps and discards its own approaches. The answer arrives after twenty seconds instead of two and is considerably more reliable on multi-step tasks.
A closer look
Such models produce a longer derivation before the actual answer and are trained so that this derivation leads to the correct result. The gain shows on multi-step tasks — mathematics, programming, planning — and is paid for with compute time.
Why it matters
The trade-off is openly visible: more compute time against fewer errors. That is exactly why choosing a model is not a decision of principle but one per task. Using a reasoning model for master data maintenance burns money and waiting time; leaving it out of a multi-step check produces errors that someone later finds by hand. Anyone who needs both in the same workflow tiers them: the fast model for the regular case, the slow one for the doubtful case — and the rule for when to switch is a substantive decision, not a technical one.
Commonly misunderstood
The visible chain of thought is an output, not an explanation. It can look plausible and still not be the path by which the answer came about.