GitHub Copilot Fusion launched Wednesday at GitHub Universe. It decides which AI model, or mix of models, handles each step of a coding task, so the developer no longer has to. GitHub also published its yearly Octoverse report the same day, and its headline number is how much work agents now start.
What GitHub Copilot Fusion does
A GitHub blog post by Jay Parikh, Microsoft’s EVP of CoreAI, says Fusion picks a model based on what a step needs: reasoning, code generation, debugging or tool use. The unified system behind Copilot’s agents is also now in Foundry, Microsoft’s managed service for long-running agents.
Parikh said the top tier, Fusion Max, came within two points of Claude Opus 5.5 Max on the Artificial Analysis Coding Agent Index. It did so at half the cost and with 40% less time per task. Satya Nadella made a similar claim on LinkedIn: frontier-level quality 40% faster and at up to 77% lower cost. These are company figures. Nobody has independently verified them.
GitHub also introduced a Copilot app that connects agents and tools across projects. It runs agent work in parallel sessions across several repositories, and a cloud agent session can move between the app, VS Code, the command line and the web. Copilot automations can now learn from their results and propose changes that developers approve. There is also a new code review tier called Max, which GitHub says posted the highest F1 score on ReviewBench for high-severity issues.
Pricing and availability
From Wednesday, GitHub Enterprise customers can use every Copilot feature without a separate subscription. They pay for the AI credits they use instead. For now the offer goes through GitHub’s sales team, and self-serve access is listed as “coming soon.” Existing Copilot Business and Copilot Enterprise customers can keep their plans or switch whenever they like.
The Octoverse numbers
The Octoverse 2026 report covers the 12 months to Sept. 30. GitHub says developers pushed code 17.9 billion times, more than in 2024 and 2025 combined. Commits reached 34.3 billion and merged pull requests 1.3 billion. By September, more than 40% of pull requests began as an agent-created draft, up from 17% in March. That works out to 110 million in September alone.
GitHub calls that agent count a floor, because it cannot detect every agent. The platform now has more than 225 million developers, 38.3 million of them in the U.S. and 32 million in India. TypeScript stayed the most-used language. Rust grew fastest among the top languages and doubled in a year. In September, 61% of Copilot users who picked their own model used more than one. That is the habit GitHub Copilot Fusion is meant to automate.
What to watch
The routing idea is sound, but the proof is cost and quality on real repositories, not a vendor benchmark. Teams that already mix models by hand should test Fusion on their own pull requests before they move contracts. The shift to credits changes budgeting too. Spend now tracks usage, so a busy agent fleet can cost more than a flat seat price. For more on how Microsoft is trying to govern agents inside companies, see our report on Microsoft’s agent risk scores, and for the model it compared itself against, our Anthropic hub.
GitHub says it is also moving to multi-region deployment on Azure, expected to finish early next year. This year it added 4 million CPU cores, 160 PB of high-speed storage and 50 Tb/s of network capacity. If more than 40% of pull requests now start with an agent, uptime matters more than it used to.
Editor’s take
Watch the cost per merged pull request, not the benchmark gap. GitHub Copilot Fusion is a bet that picking the model should be the platform’s job, and the bet is plausible. But the savings are vendor claims. The Octoverse figure for agent-started pull requests also says little about how many of those pull requests were good ones. Measure acceptance and rework before you celebrate.
