# AI 模型比较

在 Copilot Chat 中比较可用的 AI 模型，并为你的任务选择最佳模型。

## 比较 GitHub Copilot 的 AI 模型

GitHub Copilot 支持具有不同功能的多个 AI 模型。 你选择的模型会影响 Copilot Chat 和 Copilot 内联建议的响应质量和相关性。 某些模型提供较低的延迟，而另一些模型则提供更少的幻觉现象或在特定任务上表现更佳。 本指南可帮助你根据任务（而不仅仅是模型名称）选择最佳模型。

> \[!NOTE]
>
> * 不同的模型根据其令牌定价以不同的费率使用 AI credits 。 有关详细信息，请参阅 [GitHub Copilot 的模型和定价](/zh/enterprise-cloud@latest/copilot/reference/copilot-billing/models-and-pricing)。
> * 选择模型时选择 **“自动** ”选项会自动根据模型可用性和任务复杂性为任务选择最佳模型。  如果你使用的是付费 Copilot 计划，则当你使用 **“自动** ”选项时，你有资格获得模型成本的折扣。 请参阅“[关于 Copilotauto model selection](/zh/enterprise-cloud@latest/copilot/concepts/models/auto-model-selection)”。

### 按任务推荐的模型

使用下表快速查找合适的模型，请在以下部分中查看更多详细信息。

| 型号         | 任务领域 | 在（主要用例）方面表现出色 | 其他阅读材料 |
| ---------- | ---- | ------------- | ------ |
|            |      |               |        |
| GPT-5 mini |      |               |        |

```
                               | General-purpose coding and writing | Fast, accurate code completions and explanations        | [GPT-5 mini model card](https://cdn.openai.com/gpt-5-system-card.pdf) |
```

\|  |
\| GPT-5.3-Codex
\| Agentic software development | Agentic tasks        | [GPT-5.3-Codex model card](https://deploymentsafety.openai.com/gpt-5-3-codex) |
\|  |
\| GPT-5.4
\| Deep reasoning and debugging | Multi-step problem solving and architecture-level code analysis        | [GPT-5.4 model card](https://deploymentsafety.openai.com/gpt-5-4-thinking/introduction) |
\|  |
\| GPT-5.4 mini
\| Agentic software development | Codebase exploration and is especially effective when using grep-style tools        | [GPT-5.4 mini model card](https://deploymentsafety.openai.com/gpt-5-4-thinking/gpt-5-4-thinking.pdf) |
\|  |
\| GPT-5.5
\| Deep reasoning and debugging | Multi-step problem solving and architecture-level code analysis        | [GPT-5.5 model card](https://deploymentsafety.openai.com/gpt-5-5/gpt-5-5.pdf) |
\|  |
\| GPT-5.6 Luna
\| Fast help with simple or repetitive tasks | Quick, cost-efficient responses for smaller, faster coding tasks        | [GPT-5.6 model card](https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf) |
\|  |
\| GPT-5.6 Sol
\| Deep reasoning and debugging | Complex reasoning over large codebases and long-running agentic work        | [GPT-5.6 model card](https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf) |
\|  |
\| GPT-5.6 Terra
\| General-purpose coding and agent tasks | Balanced everyday interactive and agentic coding        | [GPT-5.6 model card](https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf) |
\|  |
\| Claude Fable 5
\| Long-horizon, autonomous coding and knowledge-work | First-attempt correctness through upfront reasoning, aggressive parallel tool batching, and proactive verification of pre-existing test state before debugging        | [Claude Fable 5 model card](https://anthropic.com/claude-fable-5-mythos-5-system-card) |
\|  |
\| Claude Fable 5.1
\| Long-horizon, autonomous coding and knowledge-work | Substantial, long-running coding tasks, including deep codebase research, feature development, and complex agentic workflows        | [Claude Fable 5.1 model card](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf) |
\|  |
\| Claude Haiku 4.5
\| Fast help with simple or repetitive tasks | Fast, reliable answers to lightweight coding questions        | [Claude Haiku 4.5 model card](https://assets.anthropic.com/m/99128ddd009bdcb/Claude-Haiku-4-5-System-Card.pdf) |
\|  |
\| Claude Opus 4.7
\| Deep reasoning and debugging | Complex problem-solving challenges, sophisticated reasoning        | [Claude Opus 4.7 model card](https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf) |
\|  |
\| Claude Opus 4.8
\| Deep reasoning and debugging | Complex problem-solving challenges, sophisticated reasoning        | [Claude Opus 4.8 model card](https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80fef830a37ea.pdf) |
\|  |
\| Claude Opus 4.8 (fast mode) (preview)
\| Deep reasoning and debugging | Complex problem-solving challenges, sophisticated reasoning        | [Claude Opus 4.8 model card](https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80fef830a37ea.pdf) |
\|  |
\| Claude Opus 5
\| Deep reasoning and debugging | Complex problem-solving challenges, sophisticated reasoning        | [Claude Opus 5 model card](https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf) |
\|  |
\| Claude Sonnet 4.5
\| General-purpose coding and agent tasks | Complex problem-solving challenges, sophisticated reasoning        | [Claude Sonnet 4.5 model card](https://assets.anthropic.com/m/12f214efcc2f457a/original/Claude-Sonnet-4-5-System-Card.pdf) |
\|  |
\| Claude Sonnet 4.6
\| General-purpose coding and agent tasks | Complex problem-solving challenges, sophisticated reasoning        | [Claude Sonnet 4.6 model card](https://www-cdn.anthropic.com/78073f739564e986ff3e28522761a7a0b4484f84.pdf) |
\|  |
\| Claude Sonnet 5
\| General-purpose coding and agent tasks | Complex problem-solving challenges, sophisticated reasoning        | [Claude Sonnet 5 model card](https://www-cdn.anthropic.com/9e6a1044980d8c4ed85669faf9c2a8342e2e9f1e/Claude%20Sonnet%205%20System%20Card.pdf) |
\|  |
\| Gemini 3.1 Pro
\| Deep reasoning and debugging | Effective and efficient edit-then-test loops with high tool precision        | [Gemini 3.1 Pro model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf) |
\|  |
\| Gemini 3.5 Flash
\| Fast help with simple or repetitive tasks | Fast, reliable answers to lightweight coding questions        | [Gemini 3.5 Flash model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-5-Flash-Model-Card.pdf) |
\|  |
\| Gemini 3.6 Flash
\| Fast help with simple or repetitive tasks | Fast, reliable answers to lightweight coding questions        | [Gemini 3.6 Flash model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-6-Flash-Model-Card.pdf) |
\|  |
\| Gemini 3.7 Flash
\| Fast help with simple or repetitive tasks | Fast, reliable answers to lightweight coding questions        | [Gemini 3.7 Flash model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-7-Flash-Model-Card.pdf) |
\|  |
\| Gemini 3.8 Flash
\| Fast help with simple or repetitive tasks | Fast, reliable answers to lightweight coding questions        | [Gemini 3.8 Flash model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-8-Flash-Model-Card.pdf) |
\|  |
\| MAI-Code-1-Flash

[^mai-code-1-flash]                                   | General-purpose coding and writing | Fast, accurate code completions and explanations        | [MAI-Code-1-Flash model card](https://aka.ms/mai-code-1-flash-modelcard) |
\|  |
\| MAI-Code-1.1-Flash

[^mai-code-1-flash]                                   | General-purpose coding and writing, image understanding | Fast code completions and explanations, instruction following, tool use        | [MAI-Code-1.1-Flash model card](https://microsoft.ai/pdf/MAI-Code-1.1-Flash-Model-Card.PDF) |
\|  |
\| Grok 4.5
\| General-purpose coding and agent tasks | Complex problem-solving challenges, sophisticated reasoning        | [Grok 4.5 model card](https://media.x.ai/v1/website/card-7f81d41b.pdf) |
\|  |
\| Grok 4.6
\| General-purpose coding and agent tasks | Complex problem-solving challenges, sophisticated reasoning        | [Grok 4.6 model card](https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf) |
\|  |
\| Qwen2.5
\| General-purpose coding and writing | Code generation, reasoning, and code repair / debugging        | [Qwen2.5 model card](https://arxiv.org/pdf/2409.12186) |
\|  |
\| Raptor mini
\| General-purpose coding and writing | Fast, accurate code completions and explanations        | Coming soon |
\|  |
\| Kimi K2.7 Code
\| General-purpose coding and agent tasks | Fast, reliable answers to lightweight coding questions        | [Kimi K2.7 Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code) |
\|  |
\| Kimi K3
[^kimi-k3]                                   | Agentic coding and long-context work | Multi-step agent tasks across large codebases        | [Kimi K3 model card](https://huggingface.co/moonshotai/Kimi-K3) |
\|                                        |

## 任务：常规用途的编码和编写

将这些模型用于需要平衡质量、速度和成本效益的常见开发任务。 如果没有特定要求，这些模型是不错的默认方案。

| 型号               | 为什么它很合适                                       |
| ---------------- | --------------------------------------------- |
| GPT-5.3-Codex    | 在复杂的工程任务（如功能、测试、调试、重构和评审）中，无需冗长指令即可交付更高质量的代码。 |
| GPT-5 mini       | 这是适合大多数编码和编写任务的可靠默认方案。 快速、准确且跨语言和框架工作良好。      |
| GPT-5.6 Terra    | 平衡日常交互式和代理编码的全能选择。                            |
| Raptor mini      | 专门用于提供快速而准确的内联建议和解释。                          |
| MAI-Code-1-Flash | 强大的指令跟踪和自适应推理使其成为日常编码任务、编写和多轮次开发工作流的可靠默认值。    |

### 何时使用这些模型

如果要执行以下操作，请使用这些模型之一：

* 编写或查看函数、短文件或代码变更。
* 生成文档、注释或摘要。
* 快速解释错误或意外行为。
* 在非英语编程环境中工作。

### 何时使用其他模型

如果要处理复杂的重构、体系结构决策或多步骤逻辑，请考虑使用[深层推理和调试](#task-deep-reasoning-and-debugging)中的模型。 若要完成更快、更简单的任务（如重复编辑或一次性代码建议），请参阅[快速帮助完成简单或重复的任务](#task-fast-help-with-simple-or-repetitive-tasks)。

## 任务：快速帮助完成简单或重复的任务

这些模型针对速度和响应能力进行了优化。 它们非常适合快速编辑、实用工具函数、语法帮助和轻型原型制作。 你将获得快速答案，而无需等待不必要的深度或较长的推理链。

### 建议的模型

| 型号               | 为什么它很合适                                           |
| ---------------- | ------------------------------------------------- |
| GPT-5.6 Luna     | 轻量级、经济高效的选项，用于更小、更快的任务。 GPT-5.6 系列中的最低成本模型。       |
| Claude Haiku 4.5 | 平衡快速响应与高质量输出。 非常适合小型任务和轻型代码解释。                    |
| MAI-Code-1-Flash | 处理具有自适应效率的快速编码任务，为简单请求保持简洁，并提供快速、准确的响应，而无需不必要的深度。 |

### 何时使用这些模型

如果要执行以下操作，请使用这些模型之一：

* 编写或编辑小型函数或实用工具代码。
* 询问快速语法或语言问题。
* 以最少的设置建立创意原型。
* 获取有关简单提示或编辑的快速反馈。

### 何时使用其他模型

如果要处理复杂的重构、体系结构决策或多步骤逻辑，请参阅[深层推理和调试](#task-deep-reasoning-and-debugging)。
有关需要更强常规用途的推理或更结构化输出的任务，请参阅[常规用途的编码和编写](#task-general-purpose-coding-and-writing)。

## 任务：深层推理和调试

这些模型专为需要分步推理、复杂决策或高上下文感知的任务而设计。 当你需要结构化分析、深思熟虑的代码生成或多文件理解时，它们就很合适。

### 建议的模型

| 型号                | 为什么它很合适                                                  |
| ----------------- | -------------------------------------------------------- |
| GPT-5 mini        | 提供深度推理和调试能力，与 GPT-5 相比，响应速度更快，资源使用率更低。 非常适合交互式会话和分步代码分析。 |
| GPT-5.5           | 擅长复杂的推理、代码分析和技术决策。                                       |
| GPT-5.6 Sol       | GPT-5.6 系列的最高推理上限。 最适合用于对大型代码库和要求苛刻且长时间运行的代理工作进行复杂的推理。   |
| Claude Sonnet 4.6 | 对 Sonnet 4.5 版本进行了改进，在压力下能够更可靠地完成任务并进行更智能的推理。            |
| Claude Opus 4.7   | Anthropic 的最强模型。 在 Claude Opus 4.6 基础上进行了改进。             |
| Gemini 3.1 Pro    | 长篇语境中的高级推理和科学或技术分析。                                      |

### 何时使用这些模型

如果要执行以下操作，请使用这些模型之一：

* 在跨多个文件的上下文中调试复杂问题。
* 重构大型或互连的代码库。
* 跨层规划功能或架构。
* 在库、模式或工作流之间进行权衡。
* 分析日志、性能数据或系统行为。

### 何时使用其他模型

若要完成快速迭代或轻型任务，请参阅[快速帮助完成简单或重复的任务](#task-fast-help-with-simple-or-repetitive-tasks)。
有关常规开发工作流或内容生成，请参阅[常规用途的编码和编写](#task-general-purpose-coding-and-writing)。

## 任务：使用视觉对象（图表、屏幕截图）

如果要询问有关屏幕截图、图表、UI 组件或其他视觉输入的问题，请使用这些模型。 这些模型支持多模态输入，非常适合前端工作或视觉对象调试。

| 型号                | 为什么它很合适                                                         |
| ----------------- | --------------------------------------------------------------- |
| GPT-5 mini        | 这是适合大多数编码和编写任务的可靠默认方案。 快速、准确，并支持视觉推理任务的多模式输入。 在不同的语言和框架中都能良好运作。 |
| Claude Sonnet 4.6 | 对 Sonnet 4.5 版本进行了改进，在压力下能够更可靠地完成任务并进行更智能的推理。                   |
| Gemini 3.1 Pro    | 深度推理和调试，非常适合复杂的代码生成、调试和研究工作流。                                   |

### 何时使用这些模型

如果要执行以下操作，请使用这些模型之一：

* 询问有关图表、屏幕截图或 UI 组件的问题。
* 获取有关视觉草稿或工作流的反馈。
* 通过视觉上下文了解前端行为。

> \[!TIP]
> 如果在不支持图像输入（如代码编辑器）的上下文中使用模型，则看不到视觉推理的优势。 可以使用 MCP 服务器间接访问视觉输入。 请参阅“[使用 MCP（模型上下文协议）服务器扩展 GitHub Copilot 对话助手](/zh/enterprise-cloud@latest/copilot/how-tos/provide-context/use-mcp-in-your-ide/extend-copilot-chat-with-mcp)”。

### 何时使用其他模型

如果任务涉及深层推理或大规模重构，请考虑使用[深层推理和调试](#task-deep-reasoning-and-debugging)中的模型。 若要完成纯文本任务或更简单的代码编辑，请参阅[快速帮助完成简单或重复的任务](#task-fast-help-with-simple-or-repetitive-tasks)。

## 特定于模型的注意事项

某些模型具有行为、限制或安全措施，在为任务选择它们之前，这些行为、限制或安全措施非常有用。

### Kimi K3

Kimi K3 专为长期上下文、多步骤编码和代理工作流而设计。 在预发布测试中，该模型在某些高风险提示上表现出了较高的风险，在拒绝涉及敏感主题的请求时与其他一些模型相比不太一致。 这些行为可能会反映模型的安全训练后和对齐方面的差异。 我们部署了额外的安全措施 GitHub Copilot ，以帮助缓解已确定的风险。 与任何模型一样，企业应根据特定的用例和要求评估模型功能、限制和安全措施。

## 后续步骤

[^mai-code-1-flash]: MAI models are continuously improving models. Performance and behavior may evolve over time as new checkpoints are released.

[^kimi-k3]: For important information about Kimi K3 behavior and safeguards, see [Model-specific considerations](#kimi-k3).

选择正确的模型可帮助你充分利用 Copilot。 如果不确定要使用哪种模型，请从常规用途选项开始 GPT-5 mini，然后根据需求进行调整。

* 有关详细的模型规格和定价，请参阅 [GitHub Copilot中支持的 AI 模型](/zh/enterprise-cloud@latest/copilot/reference/ai-models/supported-models)。
* 有关如何使用不同模型的其他示例，请参阅 [使用不同任务比较 AI 模型](/zh/enterprise-cloud@latest/copilot/tutorials/compare-ai-models)。
* 若要切换模型，请参阅 [更改 GitHub Copilot 对话助手 的 AI 模型](/zh/enterprise-cloud@latest/copilot/how-tos/use-ai-models/change-the-chat-model) 或 [更改适用于 GitHub Copilot 内联建议的 AI 模型](/zh/enterprise-cloud@latest/copilot/how-tos/use-ai-models/change-the-completion-model)。
* 若要了解如何 Copilot Chat 提供不同的 AI 模型，请参阅 [GitHub Copilot 的模型托管](/zh/enterprise-cloud@latest/copilot/reference/ai-models/model-hosting)。