Reasoning model

A reasoning model spends tokens thinking before it answers: restating the problem, trying an approach, checking it, and correcting course. It's the same token-by-token generation as any LLM, aimed at scratch work first. Most current frontier models can reason this way, and many decide per request how much thinking a task needs.

For builders, reasoning changes three things. Quality improves on tasks with real steps, like reconciling conflicting records, applying several policy rules at once, math, and code. Latency grows, because the user waits through the thinking before seeing the first word of the answer. And cost grows, because thinking is billed as output tokens even when you never see the thinking text.

The mistake is treating reasoning as always on or always off. A ticket-tagging route rarely benefits from deliberation and pays for it in speed; a refund-eligibility decision across a long account history might be worth several seconds. Measure quality, latency, and output tokens per feature at different effort levels, then choose with data.

Design the interface for the wait too: stream progress, show a status while the model works, and avoid reasoning-heavy settings on routes where users expect instant results.