How it works
Reasoning models are trained to produce a chain of thought before the final answer. Internal thinking can take 10-1000+ tokens depending on problem complexity. The trade-off: better answers on hard problems vs higher latency and cost on easy ones. April 2026 frontier models adapt their thinking budget automatically.
Example
A math agent asked 'What's 17% of 234?' might directly compute. Asked 'How many cans should I order to feed 47 people for 3 days at 2.3 cans per person per day?', the model reasons through the multiplication, rounding, and ordering buffer step by step.