Devin Fusion routes tasks between models mid-job

Cognition launched Devin Fusion in preview on June 29. The system runs two agents simultaneously: one using a frontier model (Fable 5, GPT-5.5, or Opus 4.8), the other using a lower-cost sidekick model. Each maintains its own cached context. A lightweight classifier decides mid-session which agent handles each step, timing switches during context compaction to avoid cache penalties.

On Cognition’s FrontierCode benchmark, Fusion achieves frontier-model accuracy at 35% lower cost. With Fable 5 specifically, cost drops 41%. Cognition reports that 88% of its own merged pull requests are now driven by the Fusion router in production.

The system handles context switches transparently to the user. Sub-tasks that are well-defined and lower-risk (running tests, filling in boilerplate, reading documentation) route to the sidekick; tasks requiring judgment or broad context route to the frontier model.

The pattern here

Devin Fusion names what many teams have been doing manually. The practice of sending simple steps to a cheaper model while reserving expensive tokens for hard reasoning is common but usually requires hand-coded routing logic. Fusion automates the routing using classifiers trained on task characteristics.

The practical limit is that a sidekick model still needs to handle its assigned steps correctly. When the budget model makes a mistake on a sub-task that the frontier model would have caught, the cost saving can be erased by rework.

For the broader question of which model to use at which level of a workflow, the guide on reasoning depth covers the same tradeoff, though Fusion applies it dynamically rather than at design time.