推理最佳实践

Learn when to use reasoning models and how they compare to GPT models.

了解何时使用推理模型以及它们与 GPT 模型的比较。

OpenAI offers two types of models: reasoning models (o1 and o3-mini, for example) and GPT models (like GPT-4o). These model families behave differently.

OpenAI 提供两种类型的模型：推理模型（例如 o1 和 o3-mini）和 GPT 模型（如 GPT-4o）。这些模型家族的行为不同。

This guide covers:

本指南涵盖：

The difference between our reasoning and non-reasoning GPT models
我们的推理模型与非推理GPT模型之间的区别
When to use our reasoning models
何时使用我们的推理模型
How to prompt reasoning models effectively
如何有效提示推理模型

Compared to GPT models, our o-series models excel at different tasks and require different prompts. One model family isn't better than the other—they're just different.

与GPT模型相比，我们的o系列模型在不同任务上表现出色，并且需要不同的提示。一个模型系列并不比另一个更好——它们只是不同。

We trained our o-series models (“the planners”) to think longer and harder about complex tasks, making them effective at strategizing, planning solutions to complex problems, and making decisions based on large volumes of ambiguous information. These models can also execute tasks with high accuracy and precision, making them ideal for domains that would otherwise require a human expert—like math, science, engineering, financial services, and legal services.

我们训练了我们的o系列模型（“规划者”）以更长时间和更深入地思考复杂任务，使它们在战略规划、解决复杂问题的方案以及基于大量模糊信息做出决策方面变得有效。这些模型还可以高精度和高准确性地执行任务，使它们在其他情况下需要人类专家的领域中变得理想——如数学、科学、工程、金融服务和法律服务。

On the other hand, our lower-latency, more cost-efficient GPT models (“the workhorses”) are designed for straightforward execution. An application might use o-series models to plan out the strategy to solve a problem, and use GPT models to execute specific tasks, particularly when speed and cost are more important than perfect accuracy.

另一方面，我们的低延迟、更具成本效益的 GPT 模型（“工作马”）旨在进行简单的执行。一个应用可能会使用 o 系列模型来规划解决问题的策略，并使用 GPT 模型来执行特定任务，特别是在速度和成本比完美准确性更重要的情况下。

What's most important for your use case?

对于你的用例，最重要的是什么？

Speed and cost → GPT models are faster and tend to cost less
速度和成本 → GPT 模型更快且通常成本更低
Executing well defined tasks → GPT models handle explicitly defined tasks well
执行明确定义的任...