GitHub Copilot Fusion launches with multi-model coordination

Multi-model coordination for coding

GitHub has released Copilot Fusion, a new feature designed to manage how AI models handle coding tasks. Instead of relying on a single model for every request, Fusion uses technology from Project HydraFusion to coordinate multiple models dynamically. It can automatically decide when to draft an answer, when to have another model judge that answer, and when to hand the work off to a more powerful system.

The system operates using three distinct execution patterns to optimize for both speed and accuracy. In the Single pattern, the AI completes the task using just one model to keep things simple and fast. The Cascade pattern involves a first model creating a draft that is then judged and potentially handed off to a higher-performance model if needed.

For the most complex requests, the Critique pattern allows multiple models to review a draft in parallel. The original model then revises its work based on the combined feedback from these reviewers. This approach is intended to ensure that extra computation and waiting time are only spent when a task actually requires it.

Performance and cost benchmarks

Independent testing by Artificial Analysis provides a look at how Fusion performs against other coding agents. In their evaluation, Copilot Fusion running at the Max setting scored 60 overall on the Coding Agent Index v2.0. This score was nearly identical to Claude Code with Opus 5.5 Max, which received a 62 overall score.

While the quality scores were close, the cost difference was significant in the benchmark. Copilot Fusion recorded an average task cost of $5.59, which is roughly 51.1% lower than the $11.42 average cost for Claude Code with Opus 5.5 Max. This suggests that the multi-model coordination helps find a more efficient path to the same result.

However, these numbers represent API costs in a benchmark and do not reflect what individual users pay for their subscriptions. The evaluation also shows that performance varies by specific task types; for example, Fusion took first place on Terminal-Bench 4.0 but scored lower than its competitors on the FrontierCode benchmark. This means the cost-to-quality ratio may shift depending on the specific type of development work being performed.

How the system manages resources

A major part of Fusion's efficiency comes from how it handles conversation history and data. GitHub designed the system to keep using the same model in the main conversation as much as possible to reuse cached tokens. This prevents the system from re-processing the same information multiple times, which helps keep costs down during long interactions.

When the system does need to bring in auxiliary models for tasks like review, it only provides those models with the specific context they need. This avoids wasteful input and ensures that the extra models aren't processing unnecessary data. By balancing cache utilization with smart model selection, GitHub aims to make multi-model coordination more practical for everyday use.

The system also considers cache boundaries and the complexity of the conversation before switching models. It tries to avoid making changes that would increase costs without providing a noticeable improvement in the quality of the code. These mechanisms are central to the goal of making high-reasoning capabilities more accessible without a linear increase in price.

User options and plan limits

Users can choose their preferred 'processing level' within the Copilot environment. The Standard level focuses on a balance of quality and speed for everyday work by choosing one model per task. The High and Max levels are designed for more complex tasks, with Max prioritizing the most advanced reasoning and analysis available.

Availability of these levels depends on the user's specific subscription plan and organizational policies. While Standard is available to all plans, High and Max are restricted to supported tiers. Additionally, enterprise administrators can restrict which specific models are allowed to be used by the Fusion system or disable certain models entirely.

GitHub is currently offering a 30% discount on the cost of models used in Copilot Fusion for paid plan users through November 30, 2026. It is important to note that this discount is separate from the 51.1% cost reduction seen in independent benchmarks. Users should monitor their actual AI credit consumption at different processing levels to determine the best value for their specific workflow.

Practicalities for developers

While Copilot Fusion automates a lot of the heavy lifting, it does not replace the need for human oversight. One specific technical detail is that Fusion does not have its own context window; it operates within the limits of the specific model being used at that moment. It also does not launch independent subagents, which is a different architectural approach to AI task management.

Developers should be particularly careful with how the tool handles file changes in a workspace. GitHub notes that if Fusion discards an intermediate draft, any file edits that draft already made to the workspace are not automatically undone. This means a developer might see changes in their directory from a path the AI eventually decided not to take.

Because the final state of files is not automatically guaranteed, even with multi-model review, it is essential to review the diff before committing any code. Users should verify which models are available in their specific selection candidates to understand the capabilities of their current setup. Success with the tool depends on matching the processing level to the complexity of the task while maintaining a manual review of the final output.