Reasoning models are language systems optimized to solve problems through explicit intermediate steps, rather than generating a direct response from mere next-token prediction. Before responding, they dedicate compute time to "thinking": they break down the problem, explore alternatives, and verify results. Well-known examples are the OpenAI o1/o3 series, DeepSeek-R1, or Claude in its extended reasoning modes.
Their importance lies in the performance leap they offer in tasks requiring chained logic: mathematics, programming, proofs, or complex planning. They are typically trained using reinforcement learning techniques that reward correct reasoning chains, not just the final answer. This allows them to self-correct during the process.
A relevant practical nuance:
- They consume more time and tokens, which makes each query more expensive.
- They do not always pay off for simple tasks, where a conventional model responds just as well and faster.
It is advisable to reserve them for problems where structured reasoning provides a clear advantage over the added cost.