State of AI: An Empirical 100 Trillion Token Study with OpenRouter
This study addresses the lack of systematic empirical research on real-world usage patterns of large language models (LLMs), particularly in light of their enhanced multi-step reasoning capabilities, which often yield behaviors diverging significantly from expectations. Leveraging over 100 trillion tokens of real user interaction logs from the OpenRouter platform, the work employs large-scale behavioral tracking, cohort retention analysis, and cross-model and cross-regional statistical methods across task types, geographic regions, and temporal dimensions. It reveals, for the first time, the widespread adoption of open-weight models, the dominance of role-playing applications, and key retention phenomena such as the “glass slipper effect.” Notably, creative role-playing and programming assistance substantially outpace traditional productivity tasks, and early adopters exhibit markedly higher retention—offering data-driven insights to inform the design, deployment, and evaluation of LLMs.