Reasoning Is Not the Same as Inference
In the LLM field, inference and reasoning are easy to confuse, but they mean different things. Inference means running a trained model to produce an answer. It is the counterpart of training, appears in phrases such as "inference costs" and "inference servers," and applies to every model. Anthropic's glossary also uses this word for the services that run models and return answers.
Reasoning means thinking things through step by step. The reasoning in reasoning model is this one: writing out intermediate thinking before answering and reaching the answer through steps. In short, a reasoning model thinks in steps before answering; inference just means running a model.
Slower and Costlier by the Amount of Thinking
A reasoning model writes out a long thinking process before the answer. According to Anthropic, this thinking improves performance on complex tasks such as math, coding, and analysis. Google also says the thinking process helps with coding, advanced mathematics, and data analysis.
In exchange, the tokens spent thinking are billed as output, even when the thinking text is not returned to the user. The time until an answer arrives also grows longer. Google recommends less thinking for simple tasks such as fact-finding questions and classification. Anthropic also recommends lowering the thinking setting first when you want to reduce cost or waiting time.
The Reasoning Settings on This Site's Pages
Many reasoning models let you change how much they think with a setting. Even for the same model, a higher setting means longer thinking, which tends to raise ratings but also raise prices.
The pages for choosing by use combine versions of the same model that differ only in reasoning settings into one row, using the setting with the highest value on that page. You can check the value for each setting under "Reasoning settings" in the compare table. When you want something cheaper and faster, look at the values for lower settings as well before you choose.