Two rivals’ models, one agent, half the bill
Artificial Analysis benchmarked GitHub Copilot Fusion on 28 October and found it lands two points below Claude Code running Claude Opus 5.5 on its Coding Agent Index v2.0 — at $5.59 per task against $11.42, roughly half.
GitHub describes Fusion as “a runtime orchestration system” that picks and coordinates models per task. The models it draws on belong to two competitors: Anthropic’s Claude Opus 5.5 and Claude Sonnet 5.5, and OpenAI’s GPT-6 Astra and GPT-6 Sol. Artificial Analysis tested the maximum setting in the GitHub Copilot CLI against an index that weights its three constituent benchmarks equally.
On the components, Fusion took the top score on Terminal-Bench 4.0 at 71. It managed 51 on FrontierCode, third in the chart and behind Claude Code with Opus 5.5 at 56, and 57 on SWE-Atlas-QnA-AA. Token use is where the cost difference comes from: 7.2 million tokens per task, about half what Claude Opus 5.5 consumes inside Claude Code, and roughly 30% fewer than Devin Fusion running Opus 5.5 with GPT-6 Luna — which scores 62 on the index at $8.00 per task.
Microsoft is selling the plumbing
The same day, Microsoft published the layer underneath. Tina Schuchman, corporate vice president for Foundry, and Vladimir Fedorov, GitHub’s chief technology officer, describe the Copilot harness as “the orchestration layer that gives an agent the right instructions, context, execution environment, and controls”.
GitHub’s engineering team built it, Microsoft now runs its Copilot products on it, and it is available through Microsoft Foundry with the SDK on GitHub. Microsoft manages the agent loop and the sandbox; developers bring instructions, skills, tools and their choice of model. The post says the harness supports “20+ frontier models plus bring your own key”, along with MCP, agent skills and agent plugins, and lists Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code among the products already on it.
The announcement itself contains no numbers. It credits a Rust port with performance gains without quantifying them, and links out to benchmarks rather than citing any. The measured claims in this story all come from the independent evaluator, not the vendor.
What to watch
The interesting question is whether second place at half the price is the position Microsoft wants. Routing between two suppliers’ models is cheap to say and awkward to sustain: the economics depend on Anthropic’s and OpenAI’s pricing, which Microsoft does not set, and the quality depends on models it does not train. Watch whether Fusion’s index score moves when either supplier ships a new frontier model, and whether Microsoft’s own MAI models start appearing in the routing mix.
