Woofun AI reports that Cognition has launched the Devin Fusion hybrid architecture, utilizing a "Sidekick" mechanism where a cost-effective small model handles execution tasks alongside a large model responsible for planning and review. This design, combined with dynamic routing, allows for zero-cost model upgrades and independent cached contexts to minimize overhead.

Data indicates that Devin Fusion achieves a 35% average cost reduction for GPT-5.5 and Opus 4.8 level models in FrontierCode benchmarks, rising to 41% when paired with Fable 5. While 88% of internal pull requests were driven by automatic routing, performance on complex multi-file tasks involving subjective judgment dropped significantly, with scores falling from 54 to 27.