Reasoning Model

Reasoning models are a class of AI systems designed to perform extended, deliberate thinking before generating their response. Unlike standard LLMs that generate responses token by token in a single pass, reasoning models allocate additional compute to an internal thinking process where they plan, analyze, explore different approaches, and verify their logic. This "thinking" phase is analogous to how humans take time to work through complex problems before answering.

The o-series models that introduced this pattern have been folded into the mainline families. Today it appears as OpenAI's configurable reasoning effort on the GPT-5 series, Anthropic's extended thinking on Claude (on by default on the current models, which return a summary of the reasoning rather than the raw trace), and Google's thinking modes on Gemini. These models have demonstrated significant improvements on tasks requiring mathematical reasoning, scientific analysis, coding challenges, and strategic planning. The tradeoff is that reasoning models are slower and more expensive per query due to the additional compute spent on thinking.

Reasoning models represent an important direction in AI development because they show that model performance can be improved not just by training on more data but by allocating more compute at inference time. This "test-time compute" scaling means models can tackle harder problems by thinking longer, similar to how humans spend more time on difficult questions. For developers, reasoning models are most valuable for complex tasks where accuracy matters more than speed, such as code generation, mathematical proofs, and strategic analysis.