Reasoning model

What it is

A reasoning model, sometimes called a large reasoning model or LRM, is a language model optimised for deliberate step-by-step problem solving rather than fast single-pass answers. Training methods such as reinforcement learning on verifiable tasks encourage the model to plan, test and self-correct. Many vendors now ship reasoning as a mode or effort setting inside a general assistant rather than as a separate product.

Why it matters

AI search and assistant answers increasingly rely on reasoning models to break a vague question into sub-questions, run several searches and compare sources before writing. That changes what gets cited: content that resolves a specific sub-question cleanly can be pulled into an answer even if the page does not rank highly for the original query. It also means longer, more conversational prompts are viable for users, so the queries your content needs to satisfy get more complex.

How it works

Practitioners route tasks by difficulty: cheap fast models for classification, summarising and extraction, reasoning models for planning, code, analysis and multi-source research. In agentic and retrieval pipelines a reasoning model often acts as the planner, deciding what to retrieve and when it has enough evidence. Teams control cost by setting reasoning effort, capping thinking tokens, caching results and measuring whether the extra spend actually improves output quality.

When it applies

Use reasoning models when a task has several dependent steps, needs evidence weighed up, or has a checkable right answer. Avoid them for high-volume, latency-sensitive or trivially simple work where a smaller model is adequate.

Examples

• A research agent uses a reasoning model to split "which CRM suits a 40-person B2B team" into pricing, integration and migration sub-questions, then searches each one.
• An in-house team routes invoice data extraction to a small fast model and contract risk analysis to a reasoning model, cutting cost without losing accuracy.
• A marketer tests the same brief in fast mode and reasoning mode, and finds reasoning mode cites specification pages and comparison tables rather than homepage copy.

How it is measured

• Task accuracy on a fixed internal evaluation set, compared with a non-reasoning baseline
• Tokens and cost per completed task, including thinking tokens
• Latency and time to first useful output for each routing tier
• Share of AI answers citing your pages for multi-step or comparison prompts