GitHub introduces Copilot Fusion for automated AI coding

GitHub has made Copilot Fusion generally available, introducing an artificial intelligence orchestration system that automatically selects and coordinates multiple AI models to complete software development tasks without requiring programmers to manage individual model choices.

The Microsoft-owned platform announced the launch on 28 October, making the technology available across all GitHub Copilot subscription plans. Fusion evaluates coding requirements, selects appropriate models and determines whether tasks should be handled individually or through coordinated workflows involving several AI systems.

Developers can select Copilot Fusion through the existing model picker and choose among three processing levels: Standard, High and Max. Each offers different combinations of computational effort, model capability and resource consumption, allowing users to balance performance against expenditure.

GitHub said the technology combines its earlier automatic model selection capabilities with Project HydraFusion, an experimental orchestration framework designed to distribute complex programming assignments across specialised AI models.

Unlike conventional model selection, which assigns a single system to each request, Fusion can coordinate multiple models within one interaction. Its routing mechanism considers reasoning requirements, code generation, debugging, tool usage and model availability before determining an execution strategy.

The platform supports three principal execution patterns. Single assigns a task directly to one model, while Cascade allows an economical model to produce an initial response before a quality assessment determines whether escalation to a stronger system is necessary.

The third approach, Critique, involves one model generating an initial solution and another reviewing its accuracy and quality. The original model can subsequently revise its output, with the developer receiving a consolidated response rather than separate contributions from individual systems.

GitHub presented performance figures indicating that Fusion Max achieved a score of 60 on the Artificial Analysis Coding Agent Index, placing it within two points of Claude Code using Anthropic’s Opus 5.5 Max configuration.

The company said its configuration achieved that result at approximately half the average application programming interface cost per task, consuming roughly half as many tokens and requiring about 40% less processing time.

Those figures reflect the specified benchmark comparison rather than a guarantee of equivalent improvements across all programming assignments. Actual performance and expenditure depend on task complexity, selected effort levels and the models used during execution.

GitHub has retained usage-based billing for the new system. Customers are charged for the individual models involved in each operation at their applicable standard rates, meaning workflows requiring several models can incur different costs from requests completed by one model.

Customers subscribing to paid Copilot plans will receive a temporary 30% discount on model costs incurred through Fusion, providing an introductory pricing incentive while the company expands adoption.

Enterprise administrators retain control over which models the orchestration system can access. Organisational policies determine eligibility, while administrators can separately disable particular models for Fusion without removing their availability elsewhere within Copilot.

Administrators can also establish Fusion as the default model selection mechanism through enterprise-managed settings, potentially standardising how development teams allocate AI resources across coding workflows.

The launch extends beyond cloud-hosted models. GitHub has introduced experimental intelligent routing between supported locally installed AI models and cloud services, enabling eligible developers to use processing capacity available on their own computers.

On supported Windows devices, developers can configure Microsoft’s MAI Code 1.1 Flash model through the Windows ML provider. GitHub Copilot CLI additionally supports manually configured local models on Windows, macOS and Linux.

The local routing capability remains optional and requires specific configuration. GitHub cautions that selecting locally executed models does not make Copilot sessions completely offline because routing decisions continue to involve cloud infrastructure.

Fusion is supported in Visual Studio Code, the GitHub Copilot application, GitHub Copilot CLI and JetBrains integrated development environments. GitHub plans to extend compatibility to additional development clients.

The announcement formed part of GitHub Universe 2026, where the company also introduced broader changes to its AI-assisted development tools, including an application designed to coordinate coding agents and development activities across projects.

GitHub’s 2026 Octoverse findings showed that 63% of paying Copilot developers had used automatic model selection at least once during September, indicating substantial existing engagement with delegated model selection.

Participants in the earlier HydraFusion research preview will have their existing automatic selection and orchestration preferences transferred to corresponding Fusion effort levels. Developers accessing the generally available system must update supported clients before selecting the new option through the model picker.