AI in human-computer interaction has evolved significantly in recent years. While the narrative around AI often focuses on its potential to increase productivity—helping with tasks like coding, email writing, and summarizing documents—real-world data reveals a different story. A recent study by OpenRouter, which analyzed over 100 trillion tokens from billions of AI interactions, uncovers surprising truths about how people use AI models like ChatGPT, Claude, and others.
This comprehensive data study, conducted on OpenRouter’s multi-model AI platform, reveals unexpected trends in AI adoption and usage. With millions of developers and over 50% of usage coming from outside the United States, OpenRouter offers a unique snapshot of AI’s role in business and personal environments globally.
The Roleplay Revolution Nobody Saw Coming
Perhaps the most surprising finding: over half of open-source AI model interactions are used for roleplay and creative storytelling. Yes, you read that correctly. While AI is frequently lauded for its productivity capabilities, many users spend a substantial amount of time engaging in interactive fiction, gaming scenarios, and character-driven conversations.
Over 50% of open-source AI usage falls under this category, which far exceeds programming-related queries. “This counters an assumption that LLMs are mostly used for writing code, emails, or summaries,” the study notes. “In reality, many users engage with these models for companionship or exploration.”
This shift in AI usage signifies a massive, largely invisible use case for AI, one that challenges conventional views about how AI tools are applied. Rather than simply functioning as productivity boosters, these AI models are now considered tools for entertainment and personal interaction.
Read Also
Google AI Infrastructure Overview
Emerging Tech Trends for 2025
Alibaba’s Qwen Chatbot and AI Developments
Programming’s Meteoric Rise in AI Use
While roleplay remains dominant in open-source AI usage, programming tasks have seen a dramatic rise across all AI platforms. At the start of 2025, coding-related queries made up only 11% of total AI usage. By the year’s end, that figure had skyrocketed to over 50%. This trend highlights AI’s increasing integration into the software development process.
The complexity of programming queries has also grown. The average length of programming-related prompts has increased fourfold, from around 1,500 tokens to over 6,000. Some code-related requests even exceed 20,000 tokens—about the length of an entire codebase. Developers are no longer just asking for code snippets; they’re conducting full debugging sessions, architectural reviews, and multi-step problem solving.
AI models like Claude from Anthropic dominate this space, accounting for more than 60% of programming-related usage in 2025. As competition heats up, however, other models from Google, OpenAI, and open-source alternatives are quickly gaining ground.
The Chinese AI Surge
The study also uncovers a notable shift in the global AI landscape. Chinese AI models now account for approximately 30% of global usage, a significant increase from just 13% at the beginning of 2025. Models from companies like DeepSeek, Qwen (Alibaba), and Moonshot AI are gaining traction worldwide.
This surge in Chinese AI usage marks a major shift away from Western dominance in the AI space. Simplified Chinese has become the second-most common language for AI interactions, representing 5% of total global usage. The increase in Asian AI spending reflects a broader trend of international adoption, with markets like Singapore, Brazil, and Japan showing significant growth.
The Rise of “Agentic” AI
One of the most significant trends emerging from the study is the rise of agentic AI. Unlike traditional AI, which simply responds to questions or tasks, agentic AI models can execute multi-step tasks, call external tools, and reason across extended conversations.
In 2025, reasoning-optimized AI interactions have surged from nearly zero to over 50% of all AI interactions. This marks a shift in AI’s role, from a simple text generator to an autonomous agent capable of planning and execution. Instead of asking AI to write a function, users now ask AI to debug a codebase, identify performance bottlenecks, and implement solutions—a significant leap forward in AI’s capabilities.
The Glass Slipper Effect
Another fascinating insight from the study is the Glass Slipper Effect. This refers to the phenomenon where AI models that are the first to solve a critical problem create lasting user loyalty. When an AI model perfectly matches an unmet need—like the metaphorical “glass slipper”—users are more likely to stick with it long-term.
For example, Google’s Gemini 2.5 Pro retained 40% of users at month five, a significantly higher retention rate compared to later cohorts. This demonstrates that being the first to solve a high-value problem gives companies a durable competitive advantage, not just in terms of user adoption, but in user loyalty as well.
