Devin Fusion (8-minute read)

Engineering teams face the cost problem of using the most expensive model for every task. Existing model-mixing tools scored well on benchmarks but had limitations in actual code-writing ability.

Cognition introduced Devin Fusion, which specializes in routing between models without any degradation in intelligence. This technology maintains Frontier and Fable 5-level performance on FrontierCode, a new coding benchmark, while achieving a 35% reduction in cost.

The core architecture of Devin Fusion is the 'Sidekick' approach. This is a structure that runs two parallel agents.

• Main agent: Uses a Frontier-level model to handle core decision-making such as planning, resolving ambiguity, and final review.
• Sidekick agent: Uses a cost-efficient model to perform detailed tasks delegated by the main agent.

This Dynamic mid-session routing approach goes beyond simple benchmark optimization, ensuring robustness in real-world environments by flexibly switching models according to the difficulty and context of the task.