GitHub has introduced GitHub Copilot Fusion, a system that automatically selects and coordinates models for coding tasks, delivering benchmark performance comparable to Claude Code with Opus 5.5 Max while cutting average API costs roughly in half and completing tasks about 40% faster. The company unveiled the product Wednesday at GitHub Universe in San Francisco, positioning it as part of a broader vision to transform GitHub into what it calls "the most comprehensive software delivery system in the world." Developers can now hand off model-choice decisions to the system while keeping control over project direction and final outcomes.

GitHub reports that Copilot Fusion Max achieved a score of 60 on the Artificial Analysis Coding Agent Index, landing within two points of Claude Code with Opus 5.5 Max. The system accomplished that result using roughly half as many tokens per task and finishing each assignment in approximately 40% less time. GitHub says performance has improved since it launched Project HydraFusion as a research preview in September 2026. The new offering merges advanced runtime model orchestration from Project HydraFusion with Copilot Auto Model Selection, creating a unified system that handles both responsibilities simultaneously.

According to Hari Pokhrel, principal software engineer at AT&T, developers previously "would often choose a more powerful model for a complex task and forget to switch back for everyday activities." GitHub says developers can delegate decisions about which models to use and how they collaborate for tasks involving reasoning, code generation, debugging, or tool use. The system can strategically switch models mid-workflow, jumping between options based on what best suits each step—either running a single model for everything, implementing a "cascading" workflow where a quality check evaluates one model's response before accepting or escalating it, or deploying a "critique" workflow where one model reviews another's output. Pokhrel credits Copilot Fusion Standard with improving token efficiency at AT&T, cutting unnecessary spending, and preserving capacity for higher-value work.

The approach reflects a broader industry shift toward automated model routing. GitHub notes that Ramp recently moved model selection into Open SWE, its open-source coding agent, while Replit made its "intelligent model routing" system the default across every account over the summer. Developers using Copilot Fusion pick an effort level—Fusion Standard for routine coding, Fusion High for tougher tasks requiring deeper reasoning, or Fusion Max for the most demanding work—while the system orchestrates which models to deploy and when. Enterprise administrators can control which models Copilot Fusion selects from, disable specific models for Copilot Fusion without blocking them for the entire team, adjust policy settings for each model independently, and decide whether Copilot Fusion serves as the default for users. GitHub released Copilot Fusion alongside several other products at the conference, including the GitHub Copilot app, Copilot Projects, and Copilot Automations, all designed to reduce the mental overhead developers face when balancing performance against cost across different tasks. For organizations managing large developer teams, delegating technical orchestration decisions to an automated layer may free up strategic attention, though it also concentrates architectural discretion within a single vendor's black-box routing logic.