frontier AI labs lost the "deep research" frontier (andrewtrask.substack.com)

Recently, several "fusion models" from companies like OpenRouter, Sakana AI, and the Los Alamos National Laboratory have outperformed leading AI systems on the DRACO deep research benchmark, marking a significant shift in the AI landscape. These routed ensembles, which combine multiple AI models, have demonstrated superior accuracy at a substantially lower cost—giving users access to high-quality predictions that were previously only available from individual frontier models. This change has led to a redefinition of where users should go for the best AI performance, with routers now serving as the primary access point for deep research capabilities. The implications are profound: user power is shifting from traditional frontier AI companies to ensemble routers, enhancing interoperability and creating more competitive pricing. Economic power is similarly transitioning, as these routers can maintain quality while absorbing price fluctuations from model providers. Additionally, control over AI research has become more decentralized, as users can still access top-tier AI capabilities even if individual models are turned off. This evolution suggests that routed ensembles are now optimal for achieving a balance of high performance and low cost, potentially reshaping the future of AI development and deployment.