ChatGPT and Gemini Aren't Just for Work! Real Hot AI Use Cases Revealed: Average Input Tokens Up Nearly 4x Year-on-Year

After generative AI entered the large-scale application phase, it was widely assumed that most people would use large language models (LLMs) to enhance work efficiency—tasks like writing emails, organizing data, creating presentations, and coding could all be delegated to AI. However, a December 2025 study by OpenRouter and venture capital firm a16z revealed that real-world model usage is far more complex than imagined. By analyzing over 100 trillion tokens on the OpenRouter platform, the study found that traffic for roleplay and creative storytelling is remarkably high, programming-related demand has surged within a year, and models that solve users’ core problems early may trigger a 'Cinderella Glass Slipper Effect'.

Why Has Roleplay Become a Major Use Case? What Does OpenRouter’s Data Reveal?

OpenRouter’s research found that among the many uses of large language models, roleplay shows highly concentrated and stable demand. The study did not directly read all complete conversations but primarily used anonymized request data, sampling about 0.25% of prompts and responses. It analyzed the first 1,000 characters using classification tools and categorized content into programming, roleplay, translation, productivity, and education.

The results showed that nearly 60% of tokens classified as roleplay belonged to gaming and roleplay scenarios, with additional categories including writing resources, interactive narratives, and adult content. This indicates that many users are not merely having casual chats with AI, but are using models as tools for character interaction, plot development, and fictional scenario generation.

The study also found that open-weight models are rapidly growing in the roleplay market. The team hypothesizes that these models typically offer higher customizability, allowing developers to tailor them to specific characters, tones, and story requirements—making them ideal for roleplay chats, interactive fiction, and long-form storytelling. However, the report does not claim that all AI usage is dominated by roleplay, but rather emphasizes that this demand has long been overshadowed by the productivity narrative.

Why Are Developers Increasingly Relying on AI? What Trends Does Programming Usage Reveal?

If roleplay showcases the entertainment and interactive potential of LLMs, programming demand reflects AI’s rapid integration into software development workflows. OpenRouter data shows that in early 2025, programming-related requests accounted for about 11% of the platform’s total token volume. By year-end, this had exceeded 50%, making it the fastest-growing usage category.

Developers’ use of AI has also evolved. Previously, they might paste a small code snippet and ask for modifications. Now, they frequently input entire codebases, error logs, technical documentation, and prior conversations, asking models to debug, analyze architecture, generate code, or rewrite modules.

The study notes that the average input length per request grew from about 1,500 tokens in early 2024 to over 6,000 tokens by the end of 2025—nearly a fourfold increase. Requests involving programming understanding, debugging, and generation often exceed 20,000 tokens. This shift indicates that large language models are evolving from one-off text generators into analytical systems capable of handling vast background data and complex tasks.

Why Do Users Become Attached to Specific Models? How Does the Glass Slipper Effect Change the AI Market?

The research team describes the phenomenon of certain models developing highly sticky user bases as the 'Cinderella Glass Slipper Effect'. When a model launches and happens to be the first to solve a group of developers’ unresolved technical, cost, or stability issues, it fits like a perfectly sized glass slipper.

For example, users who began using Gemini 2.5 Pro in June 2025 still had about 40% returning to the platform by the fifth month. Similarly, early adopters of Claude 4 Sonnet in May 2025 showed comparable retention. This 'retention rate' refers to 'active retention'—users are counted even if they return after a period of inactivity.

The study suggests that early users gradually build internal processes, data pipelines, application services, and operational habits around a specific model. Once this system is integrated, switching platforms—even as competing models catch up—incurs cost and risk. Thus, the key to the LLM market may not be simply launching first, but being the first to solve valuable, previously difficult workloads.

What Do User Behaviors Imply? How Should the Taiwan Market Interpret AI Evolution?

OpenRouter’s research shows that human demand for AI extends beyond productivity. Roleplay, interactive storytelling, and companionship scenarios generate significant traffic, while programming, tool calling, and long-context tasks are also rapidly increasing—indicating that LLMs are evolving simultaneously toward consumer interaction and professional workflows.

However, OpenRouter’s user base primarily consists of developers, API clients, and third-party app users, so the findings cannot be directly applied to general consumers in Taiwan. For Taiwanese tech firms, the key takeaway is: consumer AI services should focus on character consistency, long-form storytelling, and interactive experience quality, while enterprise products must prioritize programming capability, tool integration, long-context handling, and system stability.

AI is no longer limited to a single use. It can be both a productivity tool for work and data analysis, and a platform that supports creativity, interaction, and emotional needs. The future competitive focus will shift from how many questions a model can answer, to whether it can consistently and reliably solve users’ real-world problems in specific contexts.