Reasoning Model
An LLM trained or designed to spend extra compute on hard problems before answering.
Reasoning models internally think through a problem before producing an answer. Examples include OpenAI's o-series, Claude with extended thinking, and Google's Gemini thinking models. They're slower and cost more than standard models, but they perform meaningfully better on tasks that benefit from sustained reasoning — math, complex code, multi-step planning, scientific analysis.
For operators, the rule of thumb is: use a reasoning model when getting it right matters more than getting it fast or cheap. Use a standard model when latency or cost is the constraint and the task is straightforward.
A pragmatic pattern: route requests dynamically. Hard cases go to a reasoning model; easy cases go to a fast standard model. Many production AI products in 2026 do this transparently to manage cost.