Forget benchmarks and marketing hype. We finally have a transparent look at how humans interact with Artificial Intelligence at scale. OpenRouter recently released a massive dataset covering 100 trillion tokens of real-world usage, and the results are not what you’d expect.
The Rise of Roleplay and Reasoning
One of the most shocking revelations from the OpenRouter report is that over 50% of open-source AI usage is dedicated to roleplay. While Silicon Valley focuses on productivity and coding, a massive segment of the user base is using LLMs for creative storytelling and character interaction.
Furthermore, the shift toward reasoning models is accelerating. These models now handle approximately 50% of all traffic, suggesting that users are increasingly offloading complex …
Forget benchmarks and marketing hype. We finally have a transparent look at how humans interact with Artificial Intelligence at scale. OpenRouter recently released a massive dataset covering 100 trillion tokens of real-world usage, and the results are not what you’d expect.
The Rise of Roleplay and Reasoning
One of the most shocking revelations from the OpenRouter report is that over 50% of open-source AI usage is dedicated to roleplay. While Silicon Valley focuses on productivity and coding, a massive segment of the user base is using LLMs for creative storytelling and character interaction.
Furthermore, the shift toward reasoning models is accelerating. These models now handle approximately 50% of all traffic, suggesting that users are increasingly offloading complex logic and multi-step problem solving to AI rather than using it for simple text completion.
The Chinese Model Explosion
In just one year, Chinese models like DeepSeek and Qwen have disrupted the market. Moving from nearly zero to a 30% market share, these models are proving that the dominance of Western labs is no longer a given. Their performance-to-price ratio has made them favorites among developers and power users globally.
The "Glass Slipper" Effect
Retention in the AI space follows a unique pattern called the "Glass Slipper Effect." The data shows that once a user finds a model that fits their specific niche or workflow perfectly, they tend to stay loyal to that model indefinitely. Finding that "perfect fit" is the biggest hurdle for new LLM providers.
Conclusion
The AI landscape is shifting from general-purpose assistants to specialized usage. Whether it’s the dominance of reasoning models or the unexpected popularity of roleplay, the way we use AI is evolving faster than the benchmarks can keep up with.
What’s your "Glass Slipper" model? Let’s discuss in the comments.