Route and orchestrate models with GitHub Copilot Fusion

GitHub launched the general availability of GitHub Copilot Fusion, an automated routing and orchestration engine that dynamically directs coding tasks across multiple models. As detailed in GitHub Changelog's report, the system combines the multi-model orchestration capabilities of Project HydraFusion with fine-tuned model selection techniques to balance execution speed, accuracy, and operational spend.

This release replaces manual per-task model toggling with automated workflow coordination across supported client environments. Engineering teams, enterprise administrators, and individual developers using Visual Studio Code, JetBrains IDEs, the GitHub Copilot CLI, or the standalone Copilot desktop app can configure tiered effort profiles to match task complexity.

01Orchestration workflows across effort tiers

Copilot Fusion provides three distinct effort tiers within the client model picker: Fusion Standard, Fusion High, and Fusion Max. Fusion Standard directs general coding work to versatile models, Fusion High selects capable models for deeper reasoning, and Fusion Max assigns demanding reasoning tasks to the most powerful models available.

The routing layer executes tasks using one of three runtime structures based on prompt requirements. In a Single workflow, an individual model directly processes the request. Under a Cascade workflow, an efficient model drafts code before an automated quality gate determines whether to accept the output or escalate to a higher-capacity model.

The Critique pattern uses a multi-agent loop where a primary model drafts code, a separate secondary model conducts a review pass, and the original model revises the output against that critique. Developers set the effort tier while the system manages underlying model transitions during multi-step coding sequences.

02Artificial Analysis coding index benchmark gains

Benchmarking on the Artificial Analysis Coding Agent Index places Copilot Fusion Max at a composite score of 60. This evaluation ranks the orchestrated configuration within two points of Claude Code paired with Opus 5.5 Max.

Copilot Fusion Max achieved that benchmark output using approximately half the total tokens required by Claude Code with Opus 5.5 Max. In addition to cutting token volume, the engine cut average API cost per task by about half and decreased task execution duration by about 40 percent.

AT&T reported internal token efficiency gains during early deployment across its engineering groups. Hari Pokhrel, Principal Software Engineer at AT&T, stated that Fusion Standard improved efficiency by picking models automatically, preventing engineers from leaving high-cost reasoning models active after finishing complex work.

03Client setup and CLI execution commands

GitHub Copilot Fusion is available on all Copilot tiers across Visual Studio Code, JetBrains IDEs, GitHub Copilot CLI, and the GitHub Copilot application. Existing users can access the new orchestration engine by updating their respective client plugins and opening the model selection menu.

gh extension upgrade gh-copilot

/model copilot-fusion

gh copilot --model copilot-fusion-max

/model default copilot-fusion-standard

Selecting Copilot Fusion in the model picker exposes the Standard, High, and Max effort levels. Developers transitioning from earlier research previews will see automatic migration, as active configurations of auto, HydraFusion, and HydraFusion Max map directly to their corresponding Fusion effort tiers.

04Enterprise policies and billing structures

Enterprise administrators can configure model access rules without disabling models globally across an organization. Fusion restricts model selection to models explicitly permitted under organizational policies, allowing administrators to audit the candidate model pool and set individual policy states.

Copilot Fusion can also be enforced as the default model across organizations using Enterprise managed settings. Administrators retain the ability to disable specific models from Fusion orchestration while keeping those models available for direct selection by developers who require unmanaged access.

Standard billing accounts for each underlying model at its normal list price during execution runs. For a limited introductory period, GitHub is applying a 30 percent discount to all underlying model usage costs generated through Copilot Fusion for paid Copilot subscribers.

05Supported client environments and roadmap

Copilot Fusion is currently operational in Visual Studio Code, JetBrains development environments, the GitHub Copilot standalone application, and the GitHub Copilot CLI. The service is available across all commercial and individual GitHub Copilot subscription tiers.

Engineering teams transitioning from manual selection workflows can maintain policy governance while standardizing on automated model allocation across active repositories. GitHub confirmed that support for additional development clients is currently in development.

Feedback collection and ongoing feature improvements are managed through documentation pages and the GitHub Community portal. Future updates will expand client integration points while maintaining parity with existing enterprise administrative controls.