GitHub Copilot Fusion officially launches: benchmark scores close to Claude Code, testing costs halved

Beating AI News Flash: GitHub officially launched its multi-model orchestration system Copilot Fusion at the Universe 2026 conference. Simple tasks are routed to the appropriate model, while complex tasks can first be attempted by cheaper models, then escalated to stronger models if needed, or different models can cross-check each other.

Tests by independent evaluation firm Artificial Analysis show that Fusion Max scored 60 on comprehensive coding, close to Claude Code paired with Opus 5.5 Max at 62. Fusion Max costs an average of $5.59 per problem in API fees and takes about 36 minutes; Claude Code paired with Opus 5.5 Max requires $11.42 and about 60 minutes. Fusion reduces cost by approximately 51% and shortens time by approximately 40%.

Frequent model switching may cause cache loss, but in this evaluation, Fusion achieved a cache hit rate of 97%, compared with 95% for Claude Code. However, these results only represent specific benchmark tests and do not mean that all tasks in actual use can save half.