GitHub's Copilot Fusion Scores Within 2 Points of Claude Opus 5.5 — at Half the Cost

At its Universe 2026 conference in San Francisco, GitHub announced that Copilot Fusion is generally available — and it is not a model at all. Fusion is a router and orchestrator: developers pick "Copilot Fusion" in the model selector, and the system decides which models to use for each task and how to combine them, combining the multi-model orchestration of Project HydraFusion with Copilot's automatic model selection.

The benchmark numbers are the pitch. Fusion Max scored 60 on Artificial Analysis' Coding Agent Index (AACAI), within two points of Claude Code with Opus 5.5 Max, currently the top performer — while GitHub says Fusion used about half the average API cost per task, roughly half the tokens, and about 40 percent less time per task. Those are vendor-reported comparisons against a third-party index, but they mark the first time GitHub has quantified its routing layer against the strongest frontier model outright.

Fusion offers three orchestration modes — Single, where one model handles the task; Cascade, where an efficient model drafts and a quality check escalates to a stronger one only if needed; and Critique, where one model drafts and another reviews before revision — plus three effort tiers, Standard, High and Max. It respects each organization's model policy, so admins can exclude specific models from Fusion without touching everyone else, and it is billed per underlying model at standard rates, with a limited-time 30 percent discount on model fees for paid Copilot plans. It is available today in VS Code, the Copilot app, the Copilot CLI and JetBrains IDEs.

The second headline release is GitHub Agentic Workflows, the GitHub Next open-source project, now generally available for Free, Team and Enterprise. It lets maintainers describe automated reasoning tasks — issue triage, CI failure analysis, documentation upkeep, test writing — in Markdown files with YAML frontmatter, which compile into ordinary GitHub Actions workflows. The default engine is the Copilot CLI, but teams can bring their own coding agent: Claude Code, OpenAI Codex, Google Gemini or Pi, each with its own authentication and billing, and new model routing picks models by task type and difficulty.

The safety stack around it is substantial: agents run with read-only permissions by default, inside sandboxed containers behind an Agent Workflow Firewall, with an integrity filter on content they see, validated safe outputs, and a dedicated threat-detection job that scans proposed changes before anything is applied. The GA adds self-hosted runners via Actions Runner Controller with Docker-in-Docker, optional Cloud Hypervisor or NVX micro-VM isolation, and enclaves that stop data leaking across repositories in cross-repo workflows.

GitHub's customer claims are aggressive but unverified outside the vendor: Accenture's Ben Alton says one team lead saves about 29 hours a week, team velocity more than doubled, 91 percent of the team's commits are now AI-written, and production errors fell from roughly 11,000 a day to under 1,000 after the rollout. Treat those as one enterprise's experience, not an industry average.

The strategic read: GitHub is betting that the moat in AI coding is no longer any single model but the orchestration layer that turns expensive frontier intelligence into cheap, reliable output. If a router really can match the best model at half the cost, the models themselves become interchangeable commodities — and the platform that routes between them captures the economics. Every lab now shipping its own model router suggests they have reached the same conclusion.