AA benchmarks Devin Fusion, first multi-model coding agent on its index

Artificial Analysis added Cognition's Devin Fusion to its coding agent index, and the result gives builders a concrete data point instead of a vendor claim. Devin Fusion is a dual model setup: a frontier model handles planning and hard reasoning steps, while the cheaper SWE-2 model executes routine edits and tool calls as a sidekick. Tested with GPT-6 Astra in the lead role, the pairing costs 43 percent less and completes tasks 31 percent faster than the same setup built on Claude Fable 5.1, while its accuracy score trails only slightly behind. That gap between cost and quality is the interesting part. Most coding agents still route every step through one expensive model, even when the step is a mechanical file edit that a smaller model handles just as well. Devin Fusion's benchmark suggests teams building their own agents can send the bulk of routine work to a cheap sidekick and reserve the frontier model for planning and judgment calls, without giving up much on task success. As agent frameworks mature, this lead plus sidekick pattern looks likely to become a default architecture rather than a niche optimization, and Artificial Analysis now has a first real benchmark to weigh that tradeoff against.