Straight answer · AI
AIReviewed Nov 2026

What is a reasoning model?

Short answer

A reasoning model is an LLM trained to spend extra computation 'thinking' before it answers — generating intermediate steps — which makes it far better at problems with a verifiable right answer, like math, code, and logic. The trade-off is latency and cost: it's slower and more expensive per answer, so you use it where correctness matters more than speed.

What the 'reasoning' actually is

A standard language model produces its answer in close to a single forward pass — it predicts the response more or less directly. A reasoning model is trained to do something different: generate a chain of intermediate steps first — breaking the problem down, trying approaches, checking itself — and only then produce the final answer. In practice it spends more computation at the moment you ask (often called test-time compute) to buy a better result.

The payoff shows up specifically on problems with a verifiable structure: math, code, logic puzzles, multi-constraint planning. On these, letting the model work step by step and catch its own mistakes lifts accuracy substantially over a one-shot answer.

The trade-off: correctness for speed and cost

Thinking isn't free. A reasoning model generates far more tokens per answer and takes longer to respond, which means higher cost and higher latency. That reframes the choice as an economic one, not a quality one:

  • Use a reasoning model when a wrong answer is expensive and the user will wait — a coding agent debugging, an analyst running a hard query.
  • Skip it when the task is simple or latency-sensitive — chat, autocomplete, classification — where you'd pay more and wait longer for accuracy you didn't need.

This is also why vendors ship both kinds. Anthropic's frontier launches and OpenAI's lineup offer fast models and reasoning models side by side, because the right tool depends on the job, and routing cheap tasks to an expensive model is a quiet way to burn margin — see how to price an AI feature.

The product takeaway

A reasoning model is a dial, not an upgrade. It trades latency and cost for correctness on problems that can be reasoned toward. Point it at the hard, high-stakes, checkable tasks; route everything else to a cheaper, faster model. Matching model to task is where the economics of an AI product are won or lost.

Seen in practice

(2)

Case studies where this shows up as a real decision, not a definition.

Related questions

(3)
How is a reasoning model different from a normal LLM?
A standard LLM answers in roughly one pass; a reasoning model is trained to generate intermediate reasoning — exploring, checking, and revising — before committing to a final answer, spending more compute at inference time. The architecture is similar; the difference is that it's taught to use extra 'thinking' tokens, which raises accuracy on hard, multi-step problems.
When should a PM choose a reasoning model?
When the task has a correct answer that's expensive to get wrong and the user will tolerate a slower response — complex coding, data analysis, math, multi-constraint planning. For latency-sensitive or simple tasks (chat, classification, autocomplete), a reasoning model is overkill: you pay more and wait longer for accuracy the task didn't need.
Do reasoning models eliminate hallucination?
No. They reduce errors on problems with checkable logic because the model can catch its own missteps mid-chain, but they can still be confidently wrong, especially on questions of fact rather than reasoning. More thinking helps where the answer can be reasoned toward; it doesn't help where the model simply doesn't know.

Head-to-Head