What is a reasoning model?

What it means at work

Standard replies are quick. Reasoning modes trade speed for care: the model breaks a problem into parts, weighs options and checks itself before responding. The paper Chain of Thought Prompting Elicits Reasoning in Large Language Models shows that generating a chain of thought significantly improves the ability of large language models to perform complex reasoning. Use it for multi-step logic, tricky spreadsheets, comparing contract terms or planning; skip it for a quick rewrite.

What the sources say

• A chain of thought is a series of steps. The Chain of Thought Prompting paper defines it as a series of intermediate reasoning steps.
• Training can build it in. A DeepSeek paper shows the reasoning abilities of LLMs can be incentivized through pure reinforcement learning.
• Some assistants decide for you. The Anthropic adaptive thinking docs say a simple factual question might get a direct response with no thinking block at all. Anthropic says Claude evaluates the complexity of the input and determines whether deeper reasoning would improve the answer.
• Thinking costs tokens. Anthropic says thinking tokens are billed as output tokens once, when they are generated.

An example

You ask, ‘Given these three vendor quotes with different payment terms, which is cheapest over 24 months, and what are the assumptions?’ A reasoning mode lays out the calculation. You still verify the arithmetic against your own numbers.

A common misconception

Some think reasoning models are always right. They are usually better on hard problems, but they can reason confidently from a wrong assumption. Ask to see the steps and inputs so you can spot where a mistake entered.

Sources

• Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (arXiv): “significantly improves the ability of large language models to perform complex reasoning”
• Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (arXiv): “a chain of thought – a series of intermediate reasoning steps”
• DeepSeek-AI, “DeepSeek-R1” (arXiv): “the reasoning abilities of LLMs can be incentivized through pure reinforcement learning”
• Anthropic, Adaptive thinking: “A simple factual question might get a direct response with no thinking block at all”
• Anthropic, Adaptive thinking: “Claude evaluates the complexity of the input and determines whether deeper reasoning would improve the answer”
• Anthropic, Context windows: “Thinking tokens are billed as output tokens once … when they are generated”