Building with GitHub Copilot

The Evolution of Copilot's Underlying Models

In its earliest versions, GitHub Copilot was powered by the Codex model, a general-purpose programming model developed by OpenAI. It was trained on a mixture of public and proprietary data. This model is capable of generating code in multiple programming languages and frameworks, transpiling code from one language to another, explaining, refactoring, and debugging existing code, and even generating tests.

To make Codex faster, more efficient, and especially context-aware, the OpenAI, Azure AI, and GitHub teams worked together to fine-tune it and build upon it to create what GitHub Copilot was in its first releases.

The improved AI model behind GitHub Copilot goes beyond the previous OpenAI Codex model, offering even faster code suggestions to developers. It was developed through a collaboration between OpenAI, Microsoft Azure AI, and GitHub, and offers a 13% latency improvement over the previous model. ~ GitHub Blog

GPT-4 was the next big leap in Copilot's evolution. This smarter, more capable model could understand and generate more complex code snippets and handle more nuanced prompts and context.

Today, the GitHub code assistant still uses models from OpenAI, but not exclusively. It has expanded to include a wider array of models from different providers, including Anthropic's Claude and Google's Gemini models - a diversification that can only benefit users.

As of Q4 2025, GitHub Copilot supports a wide range of models, including but not limited to:

• Anthropic Claude Sonnet 3.5
• Anthropic Claude Sonnet 3.7
• Anthropic Claude Sonnet 3.7 Thinking
• Claude Sonnet 4