Reasoning Model: Definition and Examples

A reasoning model is a category of artificial intelligence models that incorporate an explicit reflection phase before responding. Unlike classic models that generate a direct answer token by token, reasoning models first produce an internal chain of thought—sometimes visible, sometimes hidden—in which they break down the problem, explore different avenues, and verify the coherence of their reasoning. This approach draws inspiration from techniques like Chain-of-Thought (CoT) prompting, but integrates it directly into the model's architecture or training. Iconic examples include OpenAI o1, o3, and o4-mini, DeepSeek-R1, and Claude with its extended thinking mode. These models are trained through reinforcement learning to develop autonomous reasoning strategies, enabling them to solve problems in mathematics, logic, programming, and planning with significantly higher reliability. The main trade-off of reasoning models is response time and cost: the reflection phase consumes additional tokens and increases latency. That is why they are particularly suited to tasks where accuracy trumps speed—complex code analysis, scientific problem solving, multi-criteria decision making—rather than simple conversational interactions. The emergence of reasoning models marks a turning point in generative AI: instead of simply increasing model size, we improve their ability to "think" at inference time. This paradigm, sometimes called "test-time compute scaling," paves the way for systems capable of tackling increasingly sophisticated problems.