A comprehensive new study by AI infrastructure provider OpenRouter and venture capital firm a16z offers a rare glimpse into how people actually use large language models (LLMs). By analyzing over 100 trillion tokens of anonymous user interactions, the report, titled “State of AI,” reveals that real-world AI usage is more diverse and complex than many assume, challenging common narratives about the technology’s primary applications.
One of the most significant findings is the outsized popularity of creative and entertainment-focused tasks. The study reports that creative roleplay and programming assistance are the two dominant use cases, particularly for open-source models. According to the data, roleplay, which includes storytelling and interactive character dialogues, accounts for over half of all usage on open-source models. This indicates that a large segment of users engages with AI for companionship, gaming, and creative exploration, not just for productivity tasks like writing emails or summaries.
The report also highlights the steady growth of open-source AI models. These models, whose underlying code is publicly available, have captured roughly one-third of the total token volume on the OpenRouter platform. The study notes that this growth is driven by rapid innovation, especially from Chinese developers, with models from companies like DeepSeek and Qwen gaining significant traction. While proprietary models from major labs still lead, the data suggests a healthy, multi-model ecosystem where users choose different tools for different jobs.
From simple questions to complex tasks
The nature of AI interactions is also evolving. The report identifies a clear shift away from single-question-and-answer sessions toward what it calls “agentic inference.” This means users are increasingly employing AI for multi-step, automated workflows that involve reasoning, planning, and using external tools. An example of this is an AI that can access a calendar, book a flight, and send a confirmation email all in one command. This trend is reflected in the data by the rising use of models optimized for reasoning and a sharp increase in the average length of user prompts, especially for programming tasks.
Another key concept introduced in the report is the “Cinderella ‘Glass Slipper’ effect,” which describes user retention. The analysis found that some early users of a new model show extremely high, long-term loyalty. This happens when a model is the first to perfectly solve a specific, high-value problem for a user, creating a “perfect fit” like Cinderella’s slipper. These foundational user groups tend to stick with their chosen model, even when newer alternatives become available, because of the high cost and effort of switching.
Finally, the study shows that AI usage is a global phenomenon. North America accounts for less than half of the total usage, with Asia’s share growing to over 28%. The findings underscore that real-world AI application is not a monolith but a complex landscape of diverse user needs, regional trends, and evolving technologies.
