Over the past year, artificial intelligence has been consistently portrayed as a key tool for boosting productivity, from composing emails and generating code to automatically summarizing documents, seemingly rewriting the way we work. However, a large-scale data study from OpenRouter reveals a significant discrepancy between how people actually use artificial intelligence and mainstream perception.
OpenRouter releases a report on real-world human behavior when using AI.
OpenRouter is a multi-model AI inference platform that integrates data research from over 60 vendors and more than 300 models, including open-source models such as OpenAI, Anthropic, DeepSeek, and Meta LLaMA. It analyzed anonymized metadata from over 100 trillion tokens and billions of model interactions, outlining real-world AI usage behavior globally without accessing the actual dialogue content. Notably, this research analyzed metadata from billions of interactions without accessing the actual dialogue text, thus revealing behavioral patterns while protecting user privacy.
The main use of the LLM open-source artificial language model is actually this?
Research shows that by the end of 2025, the usage of large open-source LLM language models is expected to account for about one-third of the total, with significant growth occurring with each major version release. However, what truly defies expectations is the primary use of open-source models. Data indicates that over half of open-source model interactions are not used for productivity or business applications, but rather for role-playing, interactive fiction, and creative storytelling. The most surprising finding is that over half of open-source AI models are not used to improve productivity, but rather for role-playing and story creation.
Role-playing activities have even surpassed program-assisted learning to become the largest use case. Researchers point out that this shows many users see AI as a medium for companionship, exploration, and creation, rather than simply a productivity tool. The report states that this refutes the assumption that LLM is primarily used for writing code, emails, or summaries.
AI-based correction programming is the fastest-growing application category.
In contrast, programming is the fastest-growing application category among all models. At the beginning of 2025, programming-related requests accounted for only about 10% of the total, but by the end of the year, this had exceeded 50%. The length of related prompts has also increased significantly, indicating that developers are incorporating AI into deeper levels of debugging (debugging, architecture analysis, and system-level problem-solving processes). Anthropic's Claude series once dominated the programming field, but competition from OpenAI and Google is rapidly intensifying.
Simplified Chinese has become the world's second most used AI interactive language.
The study also reveals significant changes in the global landscape. The usage share of Chinese models jumped from 13% at the beginning of the year to approximately 30%, with models such as DeepSeek, Alibaba's Qwen, and Moonshot AI rapidly emerging. Simplified Chinese has become the second most commonly used interactive language for AI globally, and Asia's overall AI usage spending has more than doubled, with Singapore becoming the second most important user country after the United States.
AI reasoning is rapidly rising.
Another key trend is the rise of "AI reasoning." AI is no longer just answering single questions, but can engage in multi-step reasoning, invoke tools, and continuously perform tasks in long conversations. This type of interaction has grown from almost non-existent to accounting for more than half of the total within a year, symbolizing that artificial intelligence is shifting from a text generation tool to an agent system with planning and execution capabilities.
The Cinderella glass slipper effect: Solving problems first is the key to building stickiness.
The study also observed a phenomenon known as the "Cinderella Glass Shoe Effect": when a model first accurately solves a specific key need, it can build highly sticky user loyalty, far surpassing subsequent competitors. Once users deeply embed the model into their workflows, switching costs increase significantly. For example, the Google Gemini 2.5 Pro's user base in June 2025 had a retention rate of approximately 40% by the fifth month, far exceeding subsequent user groups. This challenges traditional notions about competition in artificial intelligence. While seizing the initiative is important, solving high-value problems first is the key to creating a lasting competitive advantage.
Regarding pricing, data shows that AI usage is surprisingly less sensitive to cost changes. High-priced and low-priced models coexist, and the market has not yet degenerated into a simple price competition. Quality, stability, and functional completeness can still command a premium for models.
Overall, this study paints a more complex and realistic picture of artificial intelligence. AI is not only reshaping professional work, but also changing the forms of creation, entertainment, and companionship; markets are rapidly diversifying, technology is evolving quickly, and user behavior is far more honest than marketing rhetoric. Understanding these real-world usage patterns will be key to the next stage of AI development.
This article, published on ABMedia , discusses how the real-world applications of artificial intelligence differ greatly from our expectations, revealing that the most popular use of AI is actually this.
