GitHub says Copilot Fusion now picks models for coding work

GitHub on Oct. 28 made Copilot Fusion generally available, a setting that lets its coding assistant choose which AI models to use for each task instead of leaving that choice to the developer. “Instead of picking a model for each task, choose Copilot Fusion and let Copilot intelligently route and orchestrate which models it uses for you,” GitHub said in its changelog. The company announced it at its GitHub Universe conference.

How it works: GitHub said Fusion combines its earlier Project HydraFusion, which coordinated several models at once, with Copilot’s automatic model selection. Developers pick Fusion in the model picker and then choose an effort level: Standard for everyday coding, High for challenging tasks and Max for “the most demanding tasks and complex reasoning.” GitHub said Fusion can hand a task to one model, run a “cascade” in which an efficient model drafts and a quality check decides whether to escalate to a stronger one, or run a “critique” in which a second model reviews the first model’s draft before it is revised.

The claims: GitHub said Copilot Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, “within two points of Claude Code with Opus 5.5 Max.” GitHub said it got there at about half the average API cost per task, using about half as many tokens and taking about 40% less time per task. Those are GitHub’s figures.

Who gets it: GitHub said Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. Users are billed for each model Fusion uses at that model’s standard rate, and GitHub said paid Copilot users get 30% off model costs with Fusion “for a limited time.” Administrators can limit which models Fusion draws on and make it the default. In a separate changelog, GitHub said Fusion can now choose between some on-device models and cloud models in an opt-in experimental preview, though routing decisions are still made in the cloud.

At Universe: In a post, Jay Parikh, Microsoft’s executive vice president of CoreAI, whose teams include GitHub, wrote that “developers shouldn’t have to manage every model choice to get the right result at the right price.” Parikh also announced Copilot projects, which run agent work across multiple repositories, and said GitHub Enterprise customers can now use all Copilot features without a separate subscription, paying for the AI credits they use.

Why it matters: Fusion moves the choice of model from the developer to GitHub’s router. That has three effects. For cost, GitHub bills each model Fusion calls at its standard rate, so the bill depends on routing decisions the developer no longer makes directly, which is why the administrator controls matter. For the AI labs, the router decides which companies’ models get the traffic in GitHub’s coding assistant, within the models an organization allows. For reproducibility, the same prompt may be handled by different models, or a draft-and-review chain, from one run to the next, which makes results harder to compare. GitHub’s benchmark claim that Fusion comes close to Claude Code at about half the average API cost per task is the company’s own reading of a third-party index, and what any one team pays will depend on how Fusion routes that team’s work.