GitHub Copilot Fusion Mixes Models on Every Task

GitHub has turned model choice into a setting it makes for you. At its Universe conference in San Francisco on 28 October, the company launched Copilot Fusion, which sits in the Copilot model picker but is not a model: it is a router that decides, request by request, which model or combination of models should handle the work, then hands back one response. Jay Parikh, Microsoft’s executive vice president of CoreAI, announced it in a post that frames GitHub as an “end-to-end software delivery system” for developers and their agents. According to GitHub’s documentation, Fusion is generally available in Visual Studio Code, the Copilot CLI, the new GitHub Copilot app and JetBrains IDEs. It grew out of an experiment called Project HydraFusion, which routed almost 60,000 sessions before it was renamed.

Fusion runs each prompt one of three ways. In single mode one model does the job. In cascade a cheaper model drafts an answer and a judge either accepts it or escalates to a stronger model. In critique one model drafts, several others review the draft in parallel and the first model revises. There are three effort levels to choose from: Standard, which picks one model per task and is included on every plan from Free up; High; and Max, aimed at the hardest reasoning work. A fourth, Fast, is used only for inline chat in VS Code. There is no Fusion fee. Users pay each model’s normal rate for the calls it makes, which means High and Max can use more AI credits than Standard, and paid plans get 30% off those model costs until 30 November. Hovering over the effort level after a response shows which models answered. At the Standard level, Fusion can also route to models running on your own machine, Microsoft’s MAI Code 1.1 Flash on compatible Windows hardware or Qwen3.8 Flash Next through a provider such as Ollama, a feature that is still in public preview.

GitHub’s performance claim is short and relative. On the Artificial Analysis Coding Agent Index, Parikh writes, Fusion Max came within two points of Claude Opus 5.5 at its Max setting, at half the cost and with 40% less time per task. The post gives no absolute scores, so the claim cannot yet be checked against the index itself. The documentation adds some honest caveats: Fusion has no context window of its own, so each step is limited by whichever model runs it; the pool of models and the routing patterns will change over time; and if a discarded draft has already edited files, those edits are not undone automatically.

Fusion was one of several launches. The GitHub Copilot app is an agent-first workspace that connects agents and tools across projects, and Copilot projects, available in the app from launch, run agent work across several repositories in managed parallel sessions. A cloud agent session can now be started in one place and followed across the app, VS Code, the CLI and the web. Copilot automations, workflows that run in a repository on a trigger or a schedule, can now propose improvements to themselves based on past results, for a person to approve. Copilot code review gains a Max tier above Lite and Balanced, which GitHub says scored the highest F1 on ReviewBench for high-severity issues, again without a number. The pricing change may matter more than any of these: GitHub Enterprise customers can now use every Copilot feature without a separate Copilot subscription and pay only for the AI credits they consume, the system BitsMinds explained in June. It is available through GitHub’s sales team, with self-serve “coming soon,” and existing Copilot Business and Enterprise customers can stay on their plans. Behind it, GitHub says it has added 4 million CPU cores, 160 PB of fast storage and 50 Tb/s of network capacity this year, and is moving to a multi-region Azure deployment due to finish early next year.

The Octoverse report, published the same day, shows why GitHub needs the capacity. Developers pushed code 17.9 billion times in the 12 months to 30 September, more than in 2024 and 2025 combined, and made 34.3 billion commits and 1.3 billion merged pull requests. More than 40% of pull requests in September started as an agent’s draft, up from 17% in March, about 110 million in that month alone, and 90% of September’s agent pull requests were merged within two weeks. GitHub now counts more than 225 million developers, 38.3 million in the United States and 32.0 million in India. TypeScript is still the most used language, and the number of Rust contributors doubled in a year.

The model-usage figures explain why Fusion exists. In September, 61% of Copilot users who picked a model by hand used more than one, paying users touched 47 distinct models, and no single model took more than 25% of requests. Developers were already acting as their own routers, and GitHub has decided it can do that job better than they can. The open question is cost visibility: when one request can fan out to several models at their full rates, a team on pay-per-credit Enterprise billing will want to see Max’s bill before it makes Max the default.