GitHub made Copilot Fusion generally available on 28 October, giving developers a way to choose how much effort a coding task warrants while Copilot selects the models and decides how they work together. GitHub announced the release at GitHub Universe, its developer conference in San Francisco, The New Stack reported.
Key points
- Copilot Fusion combines model selection with workflows that can use more than one model on a task.
- GitHub says Fusion Max scored 60 on a coding-agent benchmark, within two points of Claude Code with Opus 5.5 Max.
- Enterprise administrators can restrict Fusion’s model choices; paid-plan users receive a temporary discount on its model costs.
Copilot Fusion chooses models and workflows
Fusion combines two earlier GitHub systems: Project HydraFusion, which coordinated multiple models, and Copilot auto model selection, which chose a model for a request. Project HydraFusion entered research preview in September 2026. In Fusion, a developer selects the system in Copilot’s model picker rather than naming an individual model, and GitHub says it then determines which models to use for each prompt and whether they should work in sequence.
The routing can follow three workflows. Single sends the task to one model. Cascade starts with a response from an efficient model and checks its quality before either accepting it or passing the task to a stronger one — much as a draft might be checked before it needs another pass. In Critique, one model drafts, a second reviews, and the first revises. The choice of workflow sits alongside the choice of model, so a request can involve several model calls without a developer making each selection.
Fixing a bug could therefore involve a draft and a review without a manual switch between models, if Fusion chose Critique for the request. It could instead go directly to one model, or pass through a quality check before a stronger model took over. Those are the routes GitHub describes for coding tasks, rather than steps that would occur on every request.
Developers still set the effort level. GitHub describes Fusion Standard as suited to everyday coding, Fusion High as intended for challenging tasks and deeper reasoning, and Fusion Max as the option for demanding work. Selecting an effort level does not pick a particular model: Fusion makes that decision within the chosen setting. That division leaves the developer to judge the task while the system handles the model choices during it.
Fusion Max scores 60 on coding index
On the Artificial Analysis Coding Agent Index, a benchmark for coding agents, GitHub says Copilot Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max. GitHub also says Fusion Max reached that score in about 40% less time per task and with about half as many tokens, while its average API cost per task was about half as high. Those figures are GitHub’s account of its performance on the index.
The combination of score, cost and time is central to GitHub’s case for automated routing. Cascade can begin with an efficient model and reserve escalation for a response that fails its quality check; Critique can use another model to review work already drafted. Both arrangements give Fusion ways to allocate model calls within a task. GitHub’s published figures concern Fusion Max on the coding-agent index, while its release also offers Standard and High as separate effort choices.
Hari Pokhrel, a principal software engineer at AT&T, said Fusion Standard had improved token efficiency across the company by selecting models automatically, reducing unnecessary spending and preserving capacity for other work. He described developers choosing a more powerful model for a complex task and then forgetting to switch back for routine work. His account concerns AT&T’s use of Standard, rather than GitHub’s Fusion Max benchmark result.
GitHub gives enterprises model and billing controls
Enterprise and organisation model policies determine which models Fusion is allowed to use. GitHub says enterprise administrators can inspect that set, control each model’s policy status independently and bar a model from Fusion without disabling it for other uses. They can also make Fusion the default for Copilot users at any of its effort levels through enterprise-managed settings. The automated choice therefore operates within the models an organisation has enabled.
Fusion bills for every model it uses at that model’s standard rate, according to GitHub. Users on paid Copilot plans receive a 30% discount on model costs when using Fusion for a limited time. A task routed through more than one model can consequently involve charges for each model used, even though the developer made only one selection in the model picker.
GitHub says Fusion is available on all Copilot plans in Visual Studio Code, JetBrains IDEs, the GitHub Copilot CLI and the GitHub Copilot app. Its instructions are to update the client, choose Fusion from the model picker and then choose an effort level. For research preview participants, GitHub says earlier selections of auto, HydraFusion or HydraFusion Max automatically transition to the corresponding Fusion effort levels.
