Reasoning ModelsReasoning Models

For years, large language models answered instantly, generating each response in a single pass no matter how hard the question. In September 2024, OpenAI's o1 introduced a different mode: models that pause to reason before they reply.

These reasoning models generate a long internal chain of thought, breaking a problem into steps, exploring approaches and checking their own work before producing a final answer. The longer they are allowed to think, the better they tend to do on hard problems.

This introduced a new lever for improving AI, often called test-time compute. Instead of only making models bigger and training them longer, you could spend more computation at the moment of answering, trading time for accuracy.

The models were trained largely through reinforcement learning, rewarded for reaching correct answers on math, coding and science problems, which taught them which reasoning strategies actually pay off. The result was a sharp jump on benchmarks that demand careful, multi-step logic.

Reasoning models blurred the old line between fast, intuitive responses and slow, deliberate thought. Within months, competitors including DeepSeek's R1 followed, making step-by-step reasoning a standard capability rather than a novelty.