Large reasoning model (LRM)

A model trained to think through problems step by step before answering, spending more time and compute to handle complex tasks like math proofs, multi-file code changes, and long-horizon planning.

Standard language models answer immediately, predicting one word at a time with no pause to think through the problem. Large reasoning models (LRMs) break hard problems into intermediate steps, verify their own work, and course-correct before committing to a final answer. Examples include OpenAI's o-series models, Anthropic's Claude with extended thinking, and DeepSeek-R1. They produce better results on tasks that benefit from deliberation, at the cost of higher latency and higher token usage.

Builder example

Sending every request to a reasoning model is one of the most common ways to waste money on AI. A simple customer-service classification that runs fine on a fast, cheap model gains nothing from extended reasoning and costs five to ten times more. Route simple tasks to a fast model and reserve the reasoning model for problems where step-by-step thinking measurably improves accuracy, like complex code generation or multi-constraint planning.

A user asks how to categorize a transaction with several ambiguous facts.

Use a reasoning model, require source citations or rules, and keep a human review path for high-stakes cases.

Common confusion: Reasoning models excel at tasks that require deliberation, but they can perform worse on tasks that need direct, confident answers, like creative writing or simple factual lookups. The extra "thinking" introduces unnecessary hedging or overthinking.