In the quest for AI supremacy, leaderboard rankings and billion-parameter giants often take the spotlight. However, the true essence of artificial intelligence lies in the undercurrents of daily interactions. A revealing report from Andreessen Horowitz (a16z) and OpenRouter, released in December 2025, pulls back the curtain on this hidden world, analyzing over 100 trillion tokens from real-world LLM sessions on OpenRouter. Spanning two years, this expansive dataset is the largest empirical AI study to date, providing a mirror to human creativity and ambition.
Perhaps the most striking revelation is that more than half of all open-source model activity – around 52% – is dedicated to roleplay and creative dialogues. Users are not just engaging with AI for factual inquiries; they're crafting stories, embodying fictional personas, and exploring narratives with AI as their co-author. This creative surge emphasizes that AI's appeal is in its ability to augment human imagination, not just automate tasks.
“These roleplay interactions are the lifeblood of engagement, revealing AI's potential as a collaborative muse for storytellers and gamers alike.”
As roleplay flourishes, programming emerges as another significant force, growing from 11% to over 50% of token volume by the end of 2025. Developers are not just tinkering; they're diving deep, with prompts exceeding 20,000 tokens for tasks like code generation and debugging. Anthropic's Claude family leads this charge, processing more than 60% of all coding requests thanks to its sophisticated reasoning capabilities.
Interestingly, open-source models are gaining ground, with models like Qwen 2.5 Coder sparking a renaissance in "medium model" usage. This boom is not just quantitative but also qualitative, accelerating software development in ways reminiscent of GitHub Copilot's early days.
The rise of Chinese models is reshaping the market, with their share jumping from 1.2% to nearly 30% by the end of 2025. Labs like DeepSeek and Alibaba's Qwen lead the charge, innovating with ruthless efficiency and challenging Western dominance in areas like translation and machine learning. This shift has fragmented the market, fostering a pluralistic ecosystem where no single model dominates.
Asia's growing influence is also reflected in AI spending, doubling from 13% to 31% globally. Countries like China, Singapore, and South Korea are leading this charge, while North America's share slips below 50% for the first time. This geographic rebalancing highlights the multipolar nature of AI's future.