GitHub Reveals How Copilot Fusion Chooses Models

At GitHub Universe 2026 in San Francisco on the 28th, GitHub outlined how it developed Copilot Fusion. Megan Kozuka, GitHub’s director of product management, said the goal is to keep developers from having to decide which AI model to use. Copilot Fusion automatically selects and coordinates models according to the task, choosing among models in high demand, including OpenAI, Anthropic, Kimi, and DeepSeek, to find one suited to software development.

Kozuka said the service can reduce the burden on developers and enterprise administrators of comparing each model’s performance and cost. Companies can restrict the use of specific models and balance performance and cost among those they permit. She said the team focused on helping enterprise customers build products they can trust by ensuring unwanted models are not selected and coordinating multiple models.

The service works in three modes. Single automatically chooses one model suited to a task. Cascade connects different models across stages, such as planning and code implementation. Critic uses another model to cross-check the output.

Kozuka said the main development challenge was balancing performance and cost. The central task was to lower costs while maintaining quality at the level of state-of-the-art AI models. She said the system could lose users’ trust if it selected a model they considered low-performing. To address this, GitHub introduced tiers that let users choose high-quality processing or faster, lower-cost processing. Kozuka said the product’s ultimate value is letting users focus on development work without having to choose a model or model selector themselves.

At GitHub Universe 2026 in San Francisco on the 28th, GitHub outlined how it developed Copilot Fusion. Megan Kozuka, GitHub’s director of product management, said the goal is to keep developers from having to decide which AI model to use. Copilot Fusion automatically selects and coordinates models according to the task, choosing among models in high demand, including OpenAI, Anthropic, Kimi, and DeepSeek, to find one suited to software development.

Kozuka said the service can reduce the burden on developers and enterprise administrators of comparing each model’s performance and cost. Companies can restrict the use of specific models and balance performance and cost among those they permit. She said the team focused on helping enterprise customers build products they can trust by ensuring unwanted models are not selected and coordinating multiple models.

The service works in three modes. Single automatically chooses one model suited to a task. Cascade connects different models across stages, such as planning and code implementation. Critic uses another model to cross-check the output.

Kozuka said the main development challenge was balancing performance and cost. The central task was to lower costs while maintaining quality at the level of state-of-the-art AI models. She said the system could lose users’ trust if it selected a model they considered low-performing. To address this, GitHub introduced tiers that let users choose high-quality processing or faster, lower-cost processing. Kozuka said the product’s ultimate value is letting users focus on development work without having to choose a model or model selector themselves.