The release of the o1 reasoning model on December 5, 2024, marked a significant shift in large language models (LLMs) from basic text generation to more complex reasoning and multi-step inference. This change has accelerated the development and deployment of LLM applications, but empirical data on real-world usage has been lacking. A study conducted using over 100 trillion tokens from the OpenRouter platform sought to fill this gap, providing insights into how users interact with LLMs across various tasks and regions. The findings reveal notable trends, such as the growing adoption of open-weight models and the unexpected popularity of creative roleplay applications alongside traditional productivity tasks. The study also identified a phenomenon termed the Cinderella "Glass Slipper" effect, where early adopters of certain models displayed longer engagement than later users. This suggests that initial alignment between user needs and model capabilities fosters lasting usage patterns. The analysis categorized LLM usage by various factors, including model type (open vs. closed source), task categories (like programming and translation), geographic differences, and retention patterns. It highlighted how global usage varies, with distinct regional preferences and economic factors influencing model selection. The research relies on anonymized metadata from billions of interactions, allowing for a comprehensive look at usage dynamics without compromising user privacy. Overall, the study reveals a complex picture of LLM engagement that challenges some existing assumptions about how these models are utilized in practice.