GitHub has made its laetest coordination system for artificial intelligence generally available to users of the company’s Copilot platform. The system, called Copilot Fusion, automatically selects and routes tasks among several OpenAI and Anthropic models—including GPT-6 Astra and Claude Opus 5.5—to optimize for speed and expense. The update also includes the release of a standalone GitHub Copilot app.
In the Artificial Analysis Coding Agent Index v2.0, Copilot Fusion’s max setting cost $5.59 per task and used 7.2M tokens, roughly half the $11.42 cost and token volume of Claude Code using Opus 5.5. While the tool scored 71 on the index’s Terminal-Bench 4.0, it ranked third on the FrontierCode benchmark with a score of 51, trailing both Claude Code and Devin Fusion.