GitHub launches Copilot Fusion for multi-model coding

GitHub made Copilot Fusion generally available on October 28, launching a Copilot service that selects and coordinates AI models at runtime for coding tasks. The service is available on all Copilot plans in Visual Studio Code, Copilot CLI, the GitHub Copilot app and JetBrains IDEs.

Fusion evaluates each prompt and can choose a different execution pattern for the next prompt in the same session. GitHub offers Standard, High and Max effort levels. The company says Standard balances cost, quality and latency for everyday work, High adds orchestration for substantial tasks, and Max prioritizes frontier capability while still considering token efficiency.

The system supports three execution patterns. Single sends a task to one model. In Cascade, an efficient model drafts a solution, then a judge accepts it or escalates the task to a stronger model. In Critique, one model produces a draft, other models review it and the original model revises it. GitHub’s documentation of Fusion says multiple critics can work independently in parallel before their feedback is combined.

GitHub says Fusion combines the multi-model orchestration developed in Project HydraFusion with Copilot’s automatic model-selection system. Administrators can restrict the eligible model pool and set Fusion and an effort level as defaults. The models available to Fusion depend on a customer’s plan and organization or enterprise policies. GitHub does not disclose the routing model, decision thresholds or per-prompt rules that select an execution pattern.

GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, compared with 62 for Claude Code using Opus 5.5 Max. The benchmark comparison lists the same scores. The company also says Fusion Max used about half as many tokens, took about 40% less time and cost about half as much per task in that benchmark comparison.

The benchmark provider lists an estimated API cost of $5.59 per task for Fusion Max and $11.42 for the Claude Code configuration, with 36.1 minutes of active agent time versus 1.0 hour. It shows about 7.3 million tokens per Fusion task and 14.5 million for Claude Code. A separate Artificial Analysis launch article displays Fusion at 7.2 million tokens, a difference the opened sources do not explain.

The index gives equal weight to FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA. Its cost figure estimates pay-per-token API use rather than consumer-plan pricing or total production deployment cost. Results also varied by component: Fusion Max led on Terminal-Bench 4.0, 71 to 63, but trailed on FrontierCode, 51 to 56, and SWE-Atlas-QnA-AA, 57 to 66.

For context on a separate approach to coordinating models in the Copilot ecosystem, DataPhoenix has covered Microsoft’s multi-model Copilot harness.

GitHub bills each underlying model invoked by Fusion at that model’s standard rate. The company says paid-plan users receive a 30% discount on Fusion model costs through November 30, 2026.