GitHub Copilot Fusion combines Anthropic and OpenAI models, nearly matching Opus 5.5 performance at half the cost

Artificial Analysis has published an analysis of Copilot Fusion, GitHub Copilot's new system that, according to GitHub, works as a runtime orchestration layer. Its job is to select and coordinate models to optimize performance, cost and latency for each task. The models it chooses from include options from Anthropic and OpenAI: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol.

The firm evaluated Copilot Fusion's "max" configuration using GitHub Copilot CLI within Artificial Analysis's Coding Agent Index v2.0. Token costs are billed according to the underlying models that end up handling each request, so the final cost depends on which model the orchestrator picks at any given moment.

The analysis's key finding is the price-performance ratio. Copilot Fusion scores 2 points below Opus 5.5 in Claude Code, but at half the cost per task: $5.59 versus $11.42. As an additional reference, Devin Fusion, which combines Opus 5.5 (max) with GPT-6 Luna (high), scores 62 points at a cost of $8.00 per task.

The index weights its three constituent benchmarks equally. On Terminal-Bench 4.0, Copilot Fusion leads with 71 points, a lead the analysis describes as slight. On FrontierCode it places third, with 51 points, behind Claude Code with Opus 5.5 (max), which scores 56, and Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high), which posts 54. On SWE-Atlas-QnA-AA it falls further behind, with 57 points, between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.

Token consumption also stands out. Copilot Fusion used 7.2 million tokens per task on the index, roughly half as many as Claude Opus 5.5 in Claude Code and 30% fewer than Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high). Fewer tokens and the distribution of tasks across different models largely explain the lower cost per task.

Overall, the result suggests that dynamic routing across models from different providers can approach frontier performance at a significantly lower cost, although the article does not detail the criteria the orchestrator uses to assign each task. The full data, with per-task costs and per-benchmark results, is available on Artificial Analysis's Coding Agent Index page.