GitHub launched Copilot Fusion, with evaluations close to Claude Code

According to Beating report, GitHub officially launched the multi-model scheduling system Copilot Fusion at the Universe 2026 conference. Simple tasks will be handled by the appropriate model; complex tasks can first be attempted by a lower-cost model, and if the results are insufficient, it can switch to a stronger model or allow different models to check each other.

Independent evaluation agency Artificial Analysis's tests show that Fusion Max's overall programming score is 60, close to Claude Code paired with Opus 5.5 Max's score of 62. Fusion Max's average API cost per question is $5.59, taking about 36 minutes; Claude Code paired with Opus 5.5 Max costs $11.42, taking about 60 minutes, reducing costs by about 51% and shortening time by about 40%.

In the evaluation, Fusion's cache hit rate was 97%, while Claude Code's was 95%. The original text also noted that frequent model switching may cause cache loss, and the above results only represent specific benchmark tests, not implying that all tasks in actual use can save half the cost.