Cognition's Fusion Cuts Devin Coding Costs by 46%

Cognition has introduced Fusion into Devin Desktop and the Devin CLI, establishing a dual-agent framework designed to optimize coding tasks. Instead of relying on a single model or a simple router, Fusion coordinates a high-performing lead model to plan and review code alongside a more economical sidekick model that executes the delegated work. The company recommends pairing its Fable 5.1 model as the lead with SWE-2 as the sidekick, though Astra is also supported as an alternative lead.

Unlike traditional routers that lose context when switching models, Fusion maintains independent context windows, prompt caches, and tools for both agents. This architecture yields significant savings, with evaluations conducted alongside Artificial Analysis and Vals AI showing cost reductions between 11% and 46% across ten benchmarks. For example, the Fable 5.1 and SWE-2 pairing maintained steady quality, keeping four benchmark scores within 1.5 points of solo runs while improving migration scores by 2.7 points. Conversely, using Astra as a lead resulted in performance drops, with Terminal-Bench falling by 5.6 points and Vals Code Migration dropping by 6.4 points, while DeepSWE and FrontierCode remained flat.

For developers, Fusion introduces a counterintuitive dynamic where utilizing a pricier model can actually lower overall session costs by reducing the need for corrections. In a test comparing Fable 5 and Opus 4.8 as leads, Fable cost twice as much per token but resulted in 9% cheaper sessions overall and higher FrontierCode scores because it delegated tasks more efficiently. Cognition manages these interactions by tuning brief specificity, pushback capabilities, and code exploration policies. Developers can deploy the system via the Devin CLI using a standard shell installation script.