GitHub Copilot Fusion automates model selection to cut API costs

GitHub has introduced Copilot Fusion, a new orchestration system designed to automate model selection for coding tasks. Announced on October 28, 2026, at GitHub Universe in San Francisco, the tool aims to reduce developer cognitive load and lower API expenses by dynamically routing work to the most efficient models.

What happened

The new system combines advanced runtime model orchestration from Project HydraFusion with Copilot Auto Model Selection. This integration allows GitHub Copilot to handle both the decision of which model to use and the coordination of how those models interact during a task. Developers no longer need to manually switch between models for different types of work, such as reasoning, code generation, debugging, or tool use.

GitHub claims that its top-tier setting, Fusion Max, delivers benchmark performance comparable to Claude Code using Opus 5.5 Max. However, it achieves this at approximately half the average API cost per task. The company also reports that Fusion Max completes tasks about 40% faster than the competitor while using roughly half as many tokens. These metrics were highlighted during the GitHub Universe keynote as part of the company’s broader vision to become the most comprehensive software delivery system globally.

Early adopters have noted practical benefits in daily workflows. Hari Pokhrel, a principal software engineer at AT&T, stated that the system improved token efficiency within his team. He noted that developers often forget to switch back from powerful, expensive models to lighter ones for routine activities, leading to unnecessary spending. Copilot Fusion addresses this by automatically adjusting model usage based on the specific requirements of each step in the workflow.

How it works

Copilot Fusion operates by analyzing the nature of a coding task and selecting the appropriate model or combination of models to execute it. Instead of a static choice, the system can employ dynamic strategies such as cascading or critique workflows. In a cascading workflow, a initial model generates a response, which is then evaluated by a quality check. If the output meets certain criteria, it is accepted; otherwise, the task is escalated to a more capable model. In a critique workflow, one model reviews and refines the work of another.

Developers retain control over the general intensity of the effort through three preset levels. Fusion Standard is designed for everyday coding tasks, balancing speed and cost. Fusion High is intended for more challenging problems requiring deeper reasoning. Fusion Max is reserved for the most complex and demanding work, leveraging the full power of the orchestration engine. This tiered approach allows teams to align resource usage with task complexity without micromanaging individual model calls.

Enterprise administrators also have granular control over the system. They can define which models are available for Copilot Fusion to select from, effectively disabling specific models for the orchestration engine while keeping them accessible for direct use if needed. Admins can adjust policy states for each model independently and determine whether Copilot Fusion serves as the default option for their organization’s developers.

Key details

• Copilot Fusion was announced on October 28, 2026, at GitHub Universe in San Francisco.
• The system integrates Project HydraFusion orchestration with Copilot Auto Model Selection.
• GitHub claims Fusion Max matches Claude Code (Opus 5.5 Max) benchmarks at half the API cost.
• Tasks processed by Fusion Max reportedly take 40% less time and use half the tokens.
• Users choose between three effort levels: Fusion Standard, Fusion High, and Fusion Max.
• Enterprise admins can restrict which models the fusion engine is allowed to route tasks to.

Why it matters

For engineering leaders and developers, the primary value proposition is the reduction of operational friction and cost. Manually selecting the right model for every sub-task is mentally taxing and prone to error. Developers often default to the most powerful model to ensure quality, inadvertently inflating costs for simple refactoring or documentation tasks. By automating this decision, Copilot Fusion allows engineers to focus on architectural direction and outcome rather than infrastructure tuning.

The cost savings reported by GitHub are significant for organizations scaling AI-assisted development. If the claim of halving API costs holds true in production environments, it directly impacts the bottom line of AI-heavy engineering teams. Furthermore, the ability to enforce model policies at the administrative level helps companies manage compliance and budget constraints without stifling developer productivity. This shift mirrors a broader industry trend where intelligent routing is becoming the standard for AI coding agents, as seen with similar moves by Ramp and Replit.

What you can do

• Evaluate your current API spend on coding assistants to establish a baseline for comparison.
• Test the three effort levels (Standard, High, Max) on representative tasks to gauge performance differences.
• Review enterprise admin settings to define which models should be available for automatic routing.
• Monitor token usage and task completion times after enabling Copilot Fusion to verify efficiency gains.
• Educate your team on the new workflow, emphasizing that they can delegate model selection while retaining oversight.
• Compare the output quality of Fusion Max against your current primary model for complex debugging scenarios.