Cognition Unveils Devin Fusion: Main AI with "Sidekick" in Tandem, Reducing Development Costs by 35%

According to Sentinel Beating monitoring, the AI programming company Cognition has released the Devin Fusion hybrid model architecture for intelligent agent programming.

The system achieves a balance between cutting-edge model performance and lower costs through two core designs: firstly, the "Sidekick" mechanism, allowing an economical small model intelligent agent to work in parallel with a cutting-edge large model intelligent agent. The large model retains the "judgment" for planning, requirement clarification, and final review, while tasks such as code exploration, testing, and formatting validation are delegated to the small model. Both maintain independent cached contexts to avoid costly cache invalidation overhead. Secondly, dynamic routing adjusts the model dynamically during a session based on task evolution and switches during context compression, enabling "zero-cost" model upgrades.

Tests have shown that in the FrontierCode benchmark test, which measures code correctness and quality, Devin Fusion maintains the performance of cutting-edge models while achieving an average 35% reduction in development costs for GPT-5.5 and Opus 4.8 level models. When used in conjunction with Fable 5, costs can be reduced by 41% (Note: Due to a U.S. government directive, access to Fable 5 was suspended on June 12, 2026, and this reduction is based on historical test data).

In internal development, 88% of the eventually merged PRs were fully driven by Fusion's automatic routing. However, when tasks heavily rely on complex subtle development intents and subjective judgments (such as involving React/Redux multi-file cross-functional development), excessive delegation leads to a significant performance drop, with scores plummeting from 54 to 27.