GitHub launches Copilot Fusion to automate AI model selection

Automating the model choice

GitHub recently introduced Copilot Fusion, a new system designed to manage model selection and orchestration for software development. Instead of developers manually picking which AI model to use for a specific task, Fusion acts as a decision-maker. It evaluates the requirements of a coding job and automatically routes it to the most appropriate tool.

The system combines technology from Project HydraFusion and Copilot Auto Model Selection to handle these complex decisions. This allows developers to delegate the technical details of model routing while keeping full responsibility for the final direction of their work. As reported by The New Stack, the goal is to make GitHub a more comprehensive software delivery system.

By automating these choices, GitHub aims to reduce the mental load on engineers. Developers often have to weigh the performance of a powerful model against its higher cost. Fusion removes this friction by making the selection process invisible to the user.

How the orchestration works

Copilot Fusion does not just pick one model and stop there; it can orchestrate multiple models to complete a single workflow. For example, it can create a "cascading" workflow where one model performs a task and another model evaluates the result to decide if it is acceptable. It can also set up a "critique" workflow where one model reviews the work produced by another.

The system can also jump between different models mid-workflow based on what is needed for each specific step. This means a single task might start with one model for initial reasoning and switch to another for final code generation. This dynamic switching helps ensure that the most efficient model is always doing the work.

To give users control over this automation, GitHub offers three distinct effort levels. Developers can choose Fusion Standard for everyday coding, Fusion High for more challenging tasks requiring deeper reasoning, and Fusion Max for the most demanding and complex work. This allows users to set the desired intensity while the system handles the underlying logistics.

Performance and cost benefits

GitHub claims that Copilot Fusion provides significant efficiency gains over manual selection or other competing tools. In benchmarks, Fusion Max delivered performance comparable to Claude Code with Opus 5.5 Max. However, it achieved these results at about half the average API cost per task.

The system also demonstrated a significant speed advantage, completing tasks in about 40% less time. It achieved this by using about half as many tokens as the competing system. These metrics suggest that the orchestration layer is successfully finding shorter, cheaper paths to the same high-quality results.

Real-world feedback from early users supports these efficiency claims. Hari Pokhrel, a principal software engineer at AT&T, noted that the system helped improve token efficiency and reduced unnecessary spending. He stated that it helped preserve capacity for higher-value work by automating the routine choices.

Enterprise controls and vision

While the system is designed to be autonomous, GitHub provides several layers of oversight for organization leaders. Enterprise administrators can control exactly which models Copilot Fusion is allowed to choose from. They can disable specific models for the fusion system without affecting the rest of the developer team.

Administrators also have the ability to adjust the policy state for each model independently. They can decide whether Copilot Fusion should be the default model for all users in the company. This balance of automation and governance ensures that large organizations can maintain security and cost standards.

This release is part of a broader trend in the industry toward intelligent model routing. Other companies like Ramp and Replit have recently moved toward making model selection a default, automated feature. GitHub's goal is to give developers more brainspace to focus on building software rather than managing AI infrastructure.