Cognition has introduced Devin Fusion, a multi-model routing system designed to reduce software engineering costs by mixing different artificial intelligence models.

The system aims to maintain high performance while lowering operational expenses by 35% on the FrontierCode benchmark. This technology addresses the high costs associated with running frontier models for every engineering task. Specifically, the system routes tasks dynamically between a primary frontier model and a more cost-effective secondary model.

How Devin Fusion Optimizes Model Routing

The architecture relies on a “sidekick” approach where two parallel agents run simultaneously. The main agent, powered by a frontier model, retains high-level decision-making such as planning and final reviews. Meanwhile, it delegates mechanical coding and testing tasks to the secondary sidekick model.

Notably, both models maintain their own persistent, cached contexts to avoid expensive cache misses. This setup allows the system to switch models dynamically during context compaction, which avoids additional latency and cost penalties.

Performance and Cost Metrics

According to testing on the FrontierCode benchmark, the system achieved a 35% cost reduction compared to using frontier models like GPT-5.5 and Opus 4.8 alone. Furthermore, when integrated with Fable 5, the cost reduction reached 41% while maintaining equivalent performance levels.

However, access to Fable 5 was suspended on June 12, 2026, following a United States government directive. Consequently, the Fable 5 performance metrics reflect data collected prior to the suspension of access.

Internal Validation and Software Integration

Cognition tested the routing system internally with its own engineering teams. The company stated that 88% of merged pull requests during the trial period were driven entirely by the automated router.

This integration helps developers avoid using highly capable, expensive models for simple tasks like minor UI adjustments. Instead, the system allocates resources based on the specific complexity of each software development task.

Future Outlook for Multi-Model Systems

The development of Devin Fusion highlights a broader shift toward hybrid-model architectures in software development. As specialized open-source and proprietary models emerge, managing multiple models will become increasingly common.

Currently, the new routing system is available in preview for users of the company’s cloud agent platform. Cognition plans to restore Fable 5 integration once access to the model is officially restored.

Source: cognition.com