KT's AutoModelRouter ranks second in global LLM benchmark

KT's proprietary AI model routing technology, ‘AutoModelRouter’, has ranked second in the overall accuracy and cost category of the ‘RouterArena’ global benchmark for Large Language Model (LLM) routers.

RouterArena, a platform developed by researchers at Rice University, evaluates AI routers based on response accuracy, cost efficiency, and robustness against input variations using approximately 8,400 queries. KT’s technology, listed as ‘KT-ModelRouter’, competed alongside commercial solutions such as Microsoft’s ‘Azure Model Router’.

The routing technology functions by analyzing the type, difficulty, and knowledge domain of a user request to automatically select the most appropriate AI model. Rather than simply choosing the highest-performing or cheapest model, the system optimizes for a balance of target quality and cost-effectiveness. For instance, simple translation tasks are directed to cost-efficient models, while complex reasoning tasks are assigned to high-performance models.

KT plans to integrate AutoModelRouter into its ‘Token Factory’ AI service platform to optimize service quality and token usage costs. The research related to this technology has also been accepted as a formal paper for the ICLR 2026 international machine learning conference.