GitHub launches Copilot Fusion to route coding tasks to optimal AI models

Automated model coordination

GitHub has introduced Copilot Fusion, a new system designed to manage how AI models handle programming tasks. Instead of developers having to manually choose which AI model to use for a specific piece of code, Fusion acts as an autonomous router. It evaluates every incoming request to determine what it needs, such as code generation, debugging, or specific reasoning capabilities.

The system looks at the requirements of the task alongside the current health of the system and the availability of various models. It then selects the most suitable approach to balance three main factors: code quality, execution speed, and cost. This allows the software to provide a definitive outcome without requiring constant human intervention from the programmer.

As reported by RS Web Solutions, this capability is now publicly available across several major platforms. Developers can access it through Visual Studio Code, the Copilot CLI, the GitHub Copilot application, and JetBrains IDEs. It is important to understand that Fusion is a coordinator rather than a standalone AI entity.

Three modes of execution

Fusion operates using three specific execution patterns to complete tasks. The first is Single mode, where a solitary AI model fulfills the entire request. This is the most straightforward method and is typically used for simpler queries that do not require extensive verification or complex reasoning.

The second method is Cascade mode, which starts with an efficient model to create a draft. A judicator then assesses that draft to decide if a more robust, powerful model is needed to engage. The third method is Critique mode, where a draft is sent to multiple models for independent evaluations. The primary model then refines the final response based on that synthesized feedback.

These patterns are not fixed and can change during a single conversation. For example, a basic question might start in Single mode, but a complex follow-up might trigger Critique mode for extra scrutiny. While Cascade and Critique methods provide higher accuracy, they generally result in longer processing times because they require multiple interactions between different models.

Tiered performance levels

GitHub has categorized the Fusion capabilities into different tiers to suit different types of work. Fusion Standard is intended for routine coding tasks that do not require heavy lifting. In contrast, Fusion High is designed for more substantial, clearly defined objectives, such as fixing intricate bugs or making changes that span across multiple files.

For the most difficult tasks that demand advanced reasoning, the system utilizes Fusion Max. On the opposite end of the spectrum, Fusion Fast is dedicated to time-sensitive inquiries within the inline chat feature of Visual Studio Code. Users do not have the ability to choose Fusion Fast directly, as it is optimized specifically for speed in that environment.

To keep things efficient, Fusion tries to keep the main conversation on a single model whenever possible. This helps to protect cached tokens and ensures that supporting models only see the specific context they need. The system also ensures that all model selections follow specific subscription parameters, data residency mandates, and FedRAMP regulations.

Costs and security limits

There are specific financial incentives for users who are already paying for the service. Those on a paid Copilot plan can get a 30% discount on the costs of using Fusion models until November 30, 2026. However, developers should be aware that using the High and Max models will deplete more AI credits than the Standard variant.

GitHub also provides important clarifications regarding the security and autonomy of the tool. Fusion works separately from subagents and does not trigger their actions; its role is strictly limited to being a coordinator. It is not an autonomous agent manager that can take independent actions beyond routing and coordinating the models provided.

Furthermore, GitHub notes that supplementary AI reviews are not a guaranteed fix for all security vulnerabilities. While previous issues like prompt injections and specific repository takeovers have been patched, users are still encouraged to review all changes. GitHub warns that rejecting a draft does not automatically undo any modifications the draft might have already made to the workspace files.