GitHub Copilot Fusion picks and orchestrates AI models for you

GitHub has made Copilot Fusion generally available, a system that selects and orchestrates AI models for coding tasks so developers no longer have to. The company announced it on Wednesday at GitHub Universe in San Francisco.

Fusion combines the runtime model orchestration from Project HydraFusion, which launched as a research preview in September 2026, with Copilot Auto Model Selection. For work involving reasoning, code generation, debugging or tool use, developers can let Copilot route each task and decide how models work together.

How the orchestration works

After analyzing a task, Fusion picks a model and then chooses a workflow. That can be a single model handling everything, a "cascading" setup in which a quality check accepts a response or escalates it, or a "critique" setup in which one model reviews another's output. The system can also switch models mid-workflow as the needs of each step change.

Developers still set the effort level: Fusion Standard for everyday coding, Fusion High for harder problems and deeper reasoning, and Fusion Max for the most demanding work. Enterprise administrators can limit which models Fusion may choose from, adjust the policy for each model independently, and decide whether Fusion is the default for Copilot users.

Benchmark and cost claims

On the Artificial Analysis Coding Agent Index, GitHub says Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max. According to the company, it did so at about half the average API cost per task, with roughly half as many tokens and about 40% less time per task. These figures come from GitHub and have not been independently verified.

Hari Pokhrel, principal software engineer at AT&T, said in a statement that developers often pick a more powerful model for a complex task and forget to switch back for routine work. He credited Fusion Standard with improving token efficiency at AT&T and preserving capacity for higher-value work.

Part of a wider push

Fusion was released alongside the GitHub Copilot app, Copilot Projects and Copilot Automations, as part of GitHub's goal of becoming the most comprehensive software delivery system in the world. Others are moving the same way: Ramp recently described building model selection into its open-source coding agent Open SWE, and Replit made intelligent model routing the default on all accounts this summer.