---
title: OpenAI GPT 系列 Tools Calling 评测
description: >-
  使用 AiPMChat 测试 OpenAI GPT 系列模型（GPT 3.5-turbo / GPT-4 /GPT-4o） 的工具调用（Function
  Calling）能力，并展现评测结果
tags:
  - Tools Calling
  - Benchmark
  - Function Calling
  - 工具调用
  - 插件
---

# OpenAI GPT Series Tool Calling

Overview of the Tool Calling capabilities of OpenAI GPT series models:

| Model | Tool Calling Support | Streaming | Parallel | Simple Instruction Score | Complex Instruction Score |
| --- | --- | --- | --- | --- | --- |
| GPT-3.5-turbo | ✅ | ✅ | ✅ | 🌟🌟🌟 | 🌟 |
| GPT-4-turbo | ✅ | ✅ | ✅ | 🌟🌟 | 🌟🌟 |
| GPT-4o | ✅ | ✅ | ✅ | 🌟🌟🌟 | 🌟🌟 |

<Callout type={'info'}>
  For testing instructions, see [Tools Calling - Evaluation Task
  Introduction](/en/docs/usage/tools-calling#evaluation-task-introduction)
</Callout>

## GPT 3.5-turbo

### Simple Instruction Call: Weather Inquiry

Test Instruction: Instruction ①

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/65901ee2-78b8-4f56-9e0d-6407c484f434" />

<Image
  alt="Tool Calling for Simple Instruction in GPT 3.5 Turbo"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/1251dfc0-d1c4-4c3d-825e-dd6205793d53"
/>

<details>
<summary>Streaming Tool Calling Raw Output:</summary>

</details>

### Complex Instruction Call: Wenshengtu

Test Instruction: Instruction ②

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/2047665f-ab22-4da7-a390-0fb4ec5a2a14" />

<Image
  alt="Tool Calling for Complex Instruction in GPT 3.5 Turbo"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/125ad028-a621-4433-b5fa-321f8fd76302"
/>

<details>
<summary>Streaming Tool Calling Raw Output:</summary>

</details>

## GPT-4 Turbo

### Simple Instruction Call: Weather Inquiry

Test Instruction: Instruction ①

Unlike GPT-3.5 Turbo, GPT-4 Turbo did not respond with "okay" when calling Tool Calling, and after multiple tests, it remained the same. Therefore, in this follow-up of a compound instruction, it is not as good as GPT-3.5 Turbo, but the remaining two capabilities are still good.

Of course, it is also possible that GPT-4 Turbo's model has more "autonomy" and believes that it does not need to output this "okay."

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/f865d91b-b84a-4258-ae09-9d1e15eeb43d" />

<Image
  alt="Tool Calling for Simple Instruction in GPT-4 Turbo"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/19298693-7a9b-4b54-9e28-c46b541b4f41"
/>

<details>
<summary>Streaming Tool Calling Raw Output:</summary>

</details>

### Complex Instruction Call: Wenshengtu

Test Instruction: Instruction ②

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/69989faf-9b98-41ec-ba51-40cc3545d8d1" />

<Image
  alt="Tool Calling for Complex Instruction in GPT-4 Turbo"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/8329c1b2-5e36-4457-946c-ce3781b05afd"
/>

<details>
<summary>Streaming Tool Calling Raw Output:</summary>

</details>

## GPT-4o

### Simple Instruction Call: Weather Inquiry

Test Instruction: Instruction ①

Similar to GPT-3.5, GPT-4o performs very well in following compound instructions in simple instruction calls.

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/c77b65ab-0854-4e1f-a25b-ff43275bd318" />

<Image
  alt="Tool Calling for Simple Instruction in GPT-4o"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/e5d6214f-f628-4064-a330-cbd7c5d474ac"
/>

<details>
    <summary>Streaming Tool Calling Raw Output:</summary>

</details>

### Complex Instruction Call: Wenshengtu

Test Instruction: Instruction ②

<Video src="https://github.com/lobehub/lobe-chat/assets/28616219/714bd86a-3b58-4941-8323-186c3fa4c6ea" />

<Image
  alt="Tool Calling for Complex Instruction in GPT-4o"
  src="https://github.com/lobehub/lobe-chat/assets/28616219/8329c1b2-5e36-4457-946c-ce3781b05afd"
/>

<details>
    <summary>Streaming Tool Calling Raw Output:</summary>

```yml

```

</details>