A new data-driven study has offered one of the clearest pictures yet of how artificial intelligence is actually being used around the world in 2025. The analysis, conducted by OpenRouter, examined more than 100 trillion tokens of metadata from billions of interactions with a wide range of large language models. The results challenge widely held assumptions about how AI is contributing to productivity and reveal emerging behaviours that could shape the next era of the technology.
OpenRouter operates as a multi-model inference platform serving more than 300 models from over 60 providers. These include leading commercial systems as well as open-source alternatives. With more than half of its total usage coming from outside the United States and millions of developers depending on the platform, the dataset offers an unusually broad view of how AI functions across countries, industries, and user types. Although the study analysed behavioural patterns only through metadata rather than actual conversation content, the scale of the sample provides strong insight into real-world user intent and model performance.
One of the most striking findings is the rapid rise of open-source AI models, which now represent roughly one-third of all usage on the platform. Much of this growth stems from heightened global adoption following major model releases in late 2024 and 2025. Yet the nature of how these open-source systems are used defies expectations. More than half of open-source model activity appears to be driven by roleplay, character interaction, and creative storytelling rather than traditional productivity tasks. These interactions extend far beyond casual conversation, with users treating models as engines for structured narrative, gaming scenarios, and immersive imaginative experiences. The study indicates that nearly 60 percent of roleplay activity involves defined fictional settings, suggesting a level of creative engagement that has gone largely unnoticed in mainstream discussions of AI.
At the same time, programming has emerged as the fastest-growing category across all AI systems. Coding-related interactions expanded from just 11 percent of total usage at the start of 2025 to more than half by the end of the year. Developers are using AI not only to request snippets of code but also to conduct lengthy debugging sessions, evaluate architecture, and carry out multi-step problem solving. The average prompt size for programming tasks has quadrupled, often exceeding thousands of tokens. Claude models from Anthropic lead this segment, although competition has intensified as alternative models improve reasoning and context-handling capabilities.
Another major shift highlighted in the study is the rapid rise of Chinese AI models. Systems developed by DeepSeek, Qwen, and Moonshot AI now represent about 30 percent of total usage, up from 13 percent in early 2025. DeepSeek alone processed more than 14 trillion tokens during the review period. English remains the dominant language for AI interactions, but simplified Chinese has become the second-most common language globally, supported by significant increases in AI spending across Asia.
The report also introduces the emerging concept of agentic inference, a category referring to models capable of executing multi-step reasoning and using external tools. These agent-like behaviours accounted for more than half of all interactions by the end of 2025, marking a substantial shift away from simple question-and-answer exchanges. Instead of issuing isolated instructions, users increasingly depend on AI to handle extended workflows by maintaining context, evaluating state, and carrying out tasks across long sessions.
Another notable observation is the so-called Glass Slipper Effect, which shows that user retention improves significantly when a model is the first to solve a previously unmet need. According to the study, early adopters of models that deliver a breakthrough capability tend to remain highly loyal, even when competitors enter the market later. This demonstrates that being first to meet a specific problem can create long-term competitive advantage within the fast-evolving AI ecosystem.
Despite assumptions that price strongly influences user adoption, the study finds AI usage to be relatively price-inelastic. A decrease in cost leads to only a marginal increase in usage, suggesting that users prioritise quality, reliability, and capability over token pricing alone. Premium models continue to attract strong demand even alongside low-cost competitors.
Overall, the findings portray an AI landscape that is more diverse, dynamic, and behaviour-driven than previously understood. Real-world usage is expanding in unexpected directions, from creative roleplay to complex programming assistance to advanced agent-like reasoning. The geographical balance of the market is shifting, and user patterns are evolving quickly as models develop new capabilities. Understanding these trends will be increasingly important as AI becomes further integrated into daily life and global industries.
