What is a Reasoning Model?

Standard LLMs generate a response token by token in a single forward pass toward the answer, which can cause them to skip necessary intermediate logic on hard problems. Reasoning models are trained (often with reinforcement learning on verified problem-solving traces) to generate an extended chain of thought first — working through sub-problems, checking their own steps, backtracking when a path doesn't pan out — before committing to a final answer. This trades additional inference time and token usage for meaningfully better accuracy on complex tasks. The distinction matters for choosing a model: reasoning models tend to excel at coding, mathematics, and multi-step planning tasks where getting the logic right matters more than speed, while faster, non-reasoning models are often the better fit for simple classification, extraction, or conversational tasks where latency and cost dominate. Many production systems route between the two based on task complexity rather than using one model for everything. Reasoning ability also underpins agentic behavior: an agent that needs to plan a multi-step task, recover from a failed tool call, or weigh competing options benefits directly from a model that reasons explicitly rather than pattern-matching to a quick answer.