Copilot Fusion delegates model choice at GitHub

GitHub launches Copilot Fusion, a model router for its development tools. The system dynamically selects the model and execution scheme based on the task, with a stated goal of reducing costs and time.

GitHub launched Copilot Fusion on October 28, 2026, at GitHub Universe in San Francisco: an intelligent model router that chooses the right model, the right effort level, and the right execution scheme for each task on behalf of the developer. The tool is available with all GitHub Copilot plans and is aimed at coding, debugging, and tooling uses. The challenge is simple to state: less time lost arbitrating between quality, latency, and cost, more consistency in AI-assisted workflows.

The developer stays in control, Copilot Fusion manages the route

Copilot Fusion is not a single model but an orchestration system. GitHub explains that it selects in real time the models and execution patterns suited to the request, based on capacity signals for reasoning, code generation, debugging, and tool usage. The user selects Fusion in the model selector, then lets the platform decide the rest. The promise is clear: no longer choosing a model for each step of a task, especially when human memory starts to do unnecessary gymnastics.

Three schemes are available depending on complexity. The single mode entrusts the entire task to a single model. The cascade mode first has a solution drafted by an efficient model, then submits it to a judge who accepts it or routes it to a more powerful model. The critique mode launches several reviewers in parallel, aggregates their feedback, and has the initial response reworked. GitHub specifies that the system can change scheme from one prompt to another, even within the same session.

Cost, latency, and model health factor into the equation

GitHub does not settle for a vaguely “intelligent” selection. Copilot Fusion takes into account system status in real time, model availability, and natural cache boundaries to avoid unnecessary costs without quality gain. In other words, the router triggers additional calls only if the task truly deserves it. A simple request does not pay for a review it did not ask for.

The effort hierarchy is also explicit. Fusion Standard targets common tasks and balances cost, quality, and latency. Fusion High is for more complex work, with additional calls justified by the scope of the project. Fusion Max targets tasks requiring the highest level of reasoning. GitHub adds a Fusion Fast mode, reserved for Visual Studio Code inline chat and not selectable by the user. In enterprise environments, administrators can limit available models and set Copilot Fusion as the default option or not.

Des chiffres qui expliquent l’intérêt de GitHub

On the ground, GitHub highlights measurable gains. According to the company, Fusion Max achieves performance comparable to Claude Code with Opus 5.5 Max on the Artificial Analysis Coding Agent Index benchmark, with a score of 60, less than two points from the cited competing tool. GitHub also reports about 50% average API cost reduction, 50% fewer tokens used, and about 40% less time per task. At AT&T, Hari Pokhrel, principal software engineer, speaks of improved “token efficiency” and better preserved capacity for higher-value work.

  • October 28, 2026: announcement at GitHub Universe, San Francisco.
  • 60: Copilot Fusion Max score on Artificial Analysis Coding Agent Index.
  • 50%: average API cost reduction claimed versus Claude Code.
  • 40%: time reduction per task announced by GitHub.
  • AT&T: “improved token efficiency,” according to Hari Pokhrel.

GitHub also indicates that Copilot Fusion can use local models in some cases, via the public preview version of local auto-routing on Windows, macOS, and Linux, with MAI Code 1.1 Flash or Qwen3.8 Flash Next installed locally. This logic aligns with the current trend: for several vendors, model choice becomes a byproduct of the workflow rather than a manual decision at each request. It’s less flashy than a chatbot demo but much more useful when you need to deliver code without wasting tokens.

The scope of this launch goes beyond GitHub. Replit has already made its intelligent routing the default behavior on all accounts, and Ramp has described a similar movement in Open SWE. In this context, GitHub accelerates a fundamental trend: model orchestration becomes a standalone product, not just an interface detail.

To read