GitHub debuts Copilot Fusion to automate model selection and orchestration

Automating the model choice

GitHub recently introduced Copilot Fusion, a new system designed to handle model selection and orchestration for software development. Instead of developers manually picking which AI model to use for a specific piece of code, the system makes those decisions automatically. This move is part of GitHub's broader goal to become a comprehensive software delivery system.

The new tool combines technologies from Project HydraFusion and Copilot Auto Model Selection. By integrating these features, Copilot Fusion can decide which models are best suited for tasks involving reasoning, debugging, or tool use. It allows developers to focus on the direction of their work rather than the technicalities of model management.

As reported by thenewstack.io, the system is intended to reduce the mental load on engineers. Developers often struggle to balance the performance of high-end models against the lower costs of smaller ones. Copilot Fusion acts as an intelligent router to navigate these trade-offs without manual intervention.

How orchestration works

Copilot Fusion does not just pick one model and stick with it for the entire project. It can strategically switch models mid-workflow depending on what the specific step requires. This means a complex reasoning task might start with one model and transition to another as the work progresses.

The system supports several different types of workflows to achieve these results. It can use a "cascading" workflow where a quality check evaluates a response and decides whether to accept it or escalate it to a different model. It can also use a "critique" workflow where one model is specifically assigned to review the work of another.

These methods ensure that the most capable model is used only when necessary. This prevents developers from inadvertently using inefficient, high-cost models for everyday activities. By automating these transitions, the system preserves capacity for higher-value work while maintaining high output quality.

Performance and cost benefits

GitHub claims that Copilot Fusion Max delivers significant improvements over competing systems. When evaluated on the Artificial Analysis Coding Agent Index, Fusion Max scored 60 points. This was two points higher than the score for Claude Code with Opus 5.5 Max, despite using fewer resources.

The performance gains are accompanied by notable cost and speed efficiencies. GitHub states that Fusion Max achieved these results while using about half the average API cost per task. It also used half as many tokens and completed tasks in about 40% less time compared to the competing model.

Hari Pokhrel, a principal software engineer at AT&T, noted that the system helped his team improve token efficiency. He stated that the tool helped "preserve capacity for higher-value work" by reducing unnecessary spend. This highlights the practical application of the tool for large enterprises managing high volumes of code.

User and admin controls

While the system automates many decisions, it still provides developers with clear choices regarding effort levels. Users can select Fusion Standard for everyday coding tasks, Fusion High for more challenging work requiring deeper reasoning, or Fusion Max for the most demanding projects. This allows the user to set the intensity of the AI's involvement.

Enterprise administrators also have several layers of control over how the system operates. They can choose which specific models Copilot Fusion is allowed to select from for their organization. This means certain models can be disabled for the fusion system without removing them from the entire developer team.

Administrators can also independently adjust the policy state for each model. They have the ability to set whether Copilot Fusion serves as the default model for all users in the company. These controls ensure that organizations can maintain security and budget standards while still benefiting from automated orchestration.

The trend toward routing

The release of Copilot Fusion follows a broader industry trend toward intelligent model routing. Other companies are beginning to move model selection into the background of their coding agents. For example, Ramp recently moved model selection into its open-source coding agent, Open SWE.

Replit also followed this path by making an intelligent model routing system the default for every account. These moves suggest that the industry is moving toward a standard where the underlying model becomes less visible to the end user. The goal is to provide a seamless experience where the best tool is always used automatically.

GitHub released Copilot Fusion alongside other products like Copilot Projects and Copilot Automations. By deferring model selection to a dedicated system, GitHub aims to give developers more brainspace. This shift allows engineers to focus on solving problems rather than managing the infrastructure of the AI tools they use.