GitHub Copilot Fusion routes tasks to models and makes them debate
GitHub killed the model picker. Copilot Fusion now routes your coding tasks to the right AI and makes models check each other's work.
Tools · Source: The Register
What happened
GitHub just launched Copilot Fusion at its 2026 Universe conference in San Francisco. The tool automatically routes your coding tasks to the right AI model. You no longer have to guess which model works best for a specific bug or feature request. The system evaluates the prompt and picks the most efficient option for you.
Fusion offers three effort levels called Standard, High, and Max. It also uses three distinct workflows to get the job done. In Single mode, one model handles the task directly. Cascade mode lets a cheaper model draft the code while a stronger model checks it and decides whether to escalate. Critique mode forces two models to collaborate, with one drafting an initial response and another reviewing it before the original model makes a final revision.
This launch comes after a brutal year for GitHub infrastructure. Automated AI agents pushed code commits to 7.4 billion in September 2026, a massive five-fold increase from the previous year. This staggering load caught the company unprepared and caused severe outages. Uptime dropped to an abysmal 84 percent in May. GitHub is now adding four million CPU cores, 160 petabytes of storage, and redesigning its architecture to handle the AI-generated slop.
Key facts
- 7.4 billion — Code commits recorded by GitHub in September 2026
- 84 percent — GitHub's 90-day availability low point on May 7
- 4 million — New CPU cores GitHub is adding to its infrastructure
- 40% — Less time per task taken by Fusion Max compared to Opus 5.5 Max
- 30% — Temporary discount on model costs when using Copilot Fusion
Why it matters
Developers can finally stop wasting time playing model roulette. You do not need to use an expensive, heavy model for simple boilerplate code. Copilot Fusion handles the cost and performance tradeoff entirely in the background. You simply set the desired effort level, and the system orchestrates the rest. This keeps engineering teams focused on shipping completed software instead of managing API calls and token limits. It also prevents developers from accidentally burning through expensive tokens on trivial tasks.
Agentic coding is actively breaking traditional software infrastructure. GitHub had to physically separate compute from storage and migrate to a multi-region Azure deployment just to survive the surge in automated commits. As developers move from writing individual lines of code to orchestrating entire fleets of AI agents, platform providers will face unprecedented strain. The companies that host our code must completely rebuild their backend systems. If they fail to adapt, they will collapse under the sheer weight of machine-generated traffic.
For builders
Stop overpaying for simple coding tasks
Developers often use expensive models for basic tasks and forget to switch back. Fusion automatically routes simple requests to cheaper models and saves the heavy hitters for complex reasoning. Companies pay less for API tokens while maintaining high output quality.
Use AI to catch high-severity bugs
GitHub introduced Copilot code review Max for deep code inspections. It posted the highest F1 score on ReviewBench for critical issues. Engineering teams can use lighter models for everyday reviews and deploy Max when a pull request needs serious scrutiny.
Prepare for enterprise policy controls
Administrators now have granular control over which models Copilot Fusion can access. They can disable specific models for Fusion without disabling them globally across the organization. Enterprise buyers get tighter governance over their AI spend and corporate data exposure.
My take
I love that GitHub is finally treating AI like a team instead of a single oracle. Making models debate and check each other is exactly how human engineers work in the real world. The model picker was always a lazy user experience crutch, and I am glad to see it die.
Original reporting: The Register. This is my rewrite and opinion.