Comparison of AI models for GitHub Copilot
GitHub Copilot supports multiple AI models with different capabilities. The model you choose affects the quality and relevance of responses by Copilot Chat and Copilot inline suggestions. Some models offer lower latency, while others offer fewer hallucinations or better performance on specific tasks. This guide helps you pick the best model based on your task, not just model names.
Note
- Different models consume AI credits at different rates based on their token pricing. For details, see Models and pricing for GitHub Copilot.
- Choosing the Auto option when selecting a model automatically selects the optimum model for your task, based on model availability and task complexity. If you are on a paid Copilot plan, you qualify for a discount on model costs when you use the Auto option. See About Copilot auto model selection.
Recommended models by task
Use this table to find a suitable model quickly, see more detail in the sections below.
Task: General-purpose coding and writing
Use these models for common development tasks that require a balance of quality, speed, and cost efficiency. These models are a good default when you don't have specific requirements.
| Model | Why it's a good fit |
|---|---|
| GPT-5.3-Codex | Delivers higher-quality code on complex engineering tasks like features, tests, debugging, refactors, and reviews without lengthy instructions. |
| GPT-5 mini | Reliable default for most coding and writing tasks. Fast, accurate, and works well across languages and frameworks. |
| GPT-5.6 Terra | Balanced all-round choice for everyday interactive and agentic coding. |
| MAI-Code-1-Flash | Strong instruction-following and adaptive reasoning make it a reliable default for everyday coding tasks, writing, and multi-turn development workflows. |
When to use these models
Use one of these models if you want to:
- Write or review functions, short files, or code diffs.
- Generate documentation, comments, or summaries.
- Explain errors or unexpected behavior quickly.
- Work in a non-English programming environment.
When to use a different model
If you're working on complex refactoring, architectural decisions, or multi-step logic, consider a model from Deep reasoning and debugging. For faster, simpler tasks like repetitive edits or one-off code suggestions, see Fast help with simple or repetitive tasks.
Task: Fast help with simple or repetitive tasks
These models are optimized for speed and responsiveness. They’re ideal for quick edits, utility functions, syntax help, and lightweight prototyping. You’ll get fast answers without waiting for unnecessary depth or long reasoning chains.
Recommended models
| Model | Why it's a good fit |
|---|---|
| GPT-5.6 Luna | Lightweight, cost-efficient option for smaller, faster tasks. The lowest-cost model in the GPT-5.6 family. |
| Claude Haiku 4.5 | Balances fast responses with quality output. Ideal for small tasks and lightweight code explanations. |
| MAI-Code-1-Flash | Handles quick coding tasks with adaptive efficiency, stays concise for simple requests and delivers fast, accurate responses without unnecessary depth. |
When to use these models
Use one of these models if you want to:
- Write or edit small functions or utility code.
- Ask quick syntax or language questions.
- Prototype ideas with minimal setup.
- Get fast feedback on simple prompts or edits.
When to use a different model
If you’re working on complex refactoring, architectural decisions, or multi-step logic, see Deep reasoning and debugging. For tasks that need stronger general-purpose reasoning or more structured output, see General-purpose coding and writing.
Task: Deep reasoning and debugging
These models are designed for tasks that require step-by-step reasoning, complex decision-making, or high-context awareness. They work well when you need structured analysis, thoughtful code generation, or multi-file understanding.
Recommended models
| Model | Why it's a good fit |
|---|---|
| GPT-5 mini | Delivers deep reasoning and debugging with faster responses and lower resource usage than GPT-5. Ideal for interactive sessions and step-by-step code analysis. |
| GPT-5.5 | Great at complex reasoning, code analysis, and technical decision-making. |
| GPT-5.6 Sol | The highest reasoning ceiling in the GPT-5.6 family. Best for complex reasoning over large codebases and demanding, long-running agentic work. |
| Claude Sonnet 4.6 | Reliable completions and smarter reasoning under pressure. |
| Claude Opus 4.7 | Anthropic’s most powerful model. Strong at deep reasoning over large, complex codebases. |
When to use these models
Use one of these models if you want to:
- Debug complex issues with context across multiple files.
- Refactor large or interconnected codebases.
- Plan features or architecture across layers.
- Weigh trade-offs between libraries, patterns, or workflows.
- Analyze logs, performance data, or system behavior.
When to use a different model
For fast iteration or lightweight tasks, see Fast help with simple or repetitive tasks. For general development workflows or content generation, see General-purpose coding and writing.
Task: Working with visuals (diagrams, screenshots)
Use these models when you want to ask questions about screenshots, diagrams, UI components, or other visual input. These models support multimodal input and are well suited for front-end work or visual debugging.
When to use these models
Use one of these models if you want to:
- Ask questions about diagrams, screenshots, or UI components.
- Get feedback on visual drafts or workflows.
- Understand front-end behavior from visual context.
Tip
If you're using a model in a context that doesn’t support image input (like a code editor), you won’t see visual reasoning benefits. You may be able to use an MCP server to get access to visual input indirectly. See Extending GitHub Copilot Chat with Model Context Protocol (MCP) servers.
When to use a different model
If your task involves deep reasoning or large-scale refactoring, consider a model from Deep reasoning and debugging. For text-only tasks or simpler code edits, see Fast help with simple or repetitive tasks.
Model-specific considerations
Some models have behaviors, limitations, or safeguards that are useful to understand before you choose them for a task.
Kimi K3
Kimi K3 is designed for long-context, multi-step coding and agentic workflows. In pre-release testing, the model exhibited elevated risk on certain higher-risk prompts and was less consistent than some other models in refusing requests involving sensitive topics. These behaviors may reflect differences in the model's safety post-training and alignment. We have deployed additional safeguards in GitHub Copilot to help mitigate the identified risks. As with any model, enterprises should evaluate model capabilities, limitations, and safeguards in light of their particular use cases and requirements.
Next steps
Choosing the right model helps you get the most out of Copilot. If you're not sure which model to use, start with a general-purpose option like GPT-5 mini, then adjust based on your needs.
- For detailed model specs and pricing, see Supported AI models in GitHub Copilot.
- For more examples of how to use different models, see Comparing AI models using different tasks.
- To switch between models, refer to Changing the AI model for GitHub Copilot Chat or Changing the AI model for GitHub Copilot inline suggestions.
- To learn how Copilot Chat serves different AI models, see Hosting of models for GitHub Copilot.
Footnotes
-
MAI models are continuously improving models. Performance and behavior may evolve over time as new checkpoints are released. ↩ ↩2
-
For important information about Kimi K3 behavior and safeguards, see Model-specific considerations. ↩