Introduction: What We Think vs. What We Do When we think about how people use Large Language Models (LLMs), a familiar picture comes to mind: productivity. We imagine developers coding faster, marketers writing emails, and analysts summarizing dense documents. This narrative of AI as a pure productivity engine is powerful, but it's only a fraction of the story. A groundbreaking new study from OpenRouter, analyzing over 100 trillion tokens of real-world usage, reveals a much more complex and surprising picture. This deep dive into how millions of people actually use AI uncovers several counter-intuitive truths. From the "killer app" no one talks about to a strange phenomenon that decides which models win in the long run, the data challenges our core assumptions about the state of AI today. actually 1. The Biggest Use Case for Open Source AI Isn't Work; It's Play. Contrary to the dominant narrative of AI as a productivity engine, the single largest use case for open-source (OSS) models is creative roleplay. The data shows that "creative roleplay" and storytelling account for more than half (about 52%) of all OSS token usage, with programming being a distant second. This is a significant revelation. It suggests a massive user base turns to open models for these experiences, as they offer greater creative freedom and are often less constrained by the commercial content filters found in proprietary models. This finding uncovers a major, underserved consumer demand for AI-driven entertainment, companionship, and interactive fiction. While the industry focuses on enterprise solutions, a silent majority is using AI to create and explore new worlds. 2. The Quiet Rise of the "AI Agent" The way people use LLMs is undergoing a fundamental shift. We are moving away from simple, single-turn, question-and-answer interactions and toward multi-step, tool-integrated workflows. In these "agentic" workflows, the model plans, reasons, and acts to complete a complex task, often without direct human intervention at every step. This shift is evident in the data: reasoning-optimized models now handle over 50% of all usage, a dramatic increase from a negligible amount at the start of 2025. This move toward reasoning is fueled by more complex workloads, as the average user's prompt has grown nearly fourfold since early 2024 to accommodate tasks like analyzing entire codebases; a trend driven primarily by programming prompts, which are 3-4 times longer than general-purpose ones. The report puts this shift in stark terms, concluding that agentic inference is rapidly becoming the new default for interacting with AI. Soon enough, if not already, agentic inference will be taking over the majority of the inference. This means the bar for all future models is being raised. Success will no longer be about simply generating plausible text, but about handling complex, stateful tasks that require planning and action. 3. Open Source Is Quietly Capturing One-Third of the AI Market While proprietary models from major labs like Anthropic and OpenAI still lead the market, open-source models have steadily grown to capture approximately one-third of all token usage by late 2025. A key driver of this expansion has been the rise of Chinese-developed OSS models from providers like Qwen and DeepSeek. These models grew from a weekly market share of just 1.2% to nearly 30% in some weeks, rapidly gaining traction with global users. This trend points to a "durable dual structure" in the AI ecosystem. The source suggests the equilibrium between proprietary and OSS models has currently stabilized at roughly 30% for open source. This balance exists because developers choose proprietary systems for high-reliability enterprise tasks, while leveraging the cost-efficiency and customization of OSS for other critical workloads. The AI landscape is not shaping up to be a winner-take-all market. Instead, it’s a complementary ecosystem where developers increasingly use a mix of both open and closed models to get the job done. 4. Why AI Isn't a Commodity (Yet) Many have predicted that AI models would quickly become a cheap commodity, with price being the main differentiator. The data, however, tells a different story. Demand for LLMs is surprisingly price-inelastic, meaning price changes have little effect on usage. According to the study, a 10% price decrease corresponds to a tiny 0.5-0.7% increase in usage. Instead of a commoditized market, the ecosystem has segmented into distinct archetypes: Premium Leaders: Models like Anthropic's Claude 4 Sonnet command high prices but still see massive usage, proving that users are willing to pay a premium for top-tier quality and reliability. Efficient Giants: Models such as Google's Gemini Flash and DeepSeek pair low cost with high volume, making them the default choice for cost-sensitive, large-scale tasks. Premium Specialists: Ultra-expensive models like OpenAI's GPT-5 Pro are used sparingly for the most high-stakes tasks, where performance is the only thing that matters and cost is secondary. The Long Tail: Models like Qwen 2 7B Instruct and IBM Granite 4.0 Micro have rock-bottom pricing but limited reach, highlighting that capability and model-market fit are critical differentiators beyond just cost. Premium Leaders: Models like Anthropic's Claude 4 Sonnet command high prices but still see massive usage, proving that users are willing to pay a premium for top-tier quality and reliability. Premium Leaders: Efficient Giants: Models such as Google's Gemini Flash and DeepSeek pair low cost with high volume, making them the default choice for cost-sensitive, large-scale tasks. Efficient Giants: Premium Specialists: Ultra-expensive models like OpenAI's GPT-5 Pro are used sparingly for the most high-stakes tasks, where performance is the only thing that matters and cost is secondary. Premium Specialists: The Long Tail: Models like Qwen 2 7B Instruct and IBM Granite 4.0 Micro have rock-bottom pricing but limited reach, highlighting that capability and model-market fit are critical differentiators beyond just cost. The Long Tail: This segmentation reveals a crucial insight about the current state of the market. The relatively flat demand elasticity suggests LLMs are not yet a commodity; many users are willing to pay a premium for quality, capabilities, or stability. For now, specialized capabilities and trusted performance often trump cost, especially for professional or mission-critical workloads. 5. Winning in AI Is Like Cinderella's "Glass Slipper" In a market where new models are released constantly, how does any single model build a lasting advantage? The study proposes a new framework called the "Glass Slipper effect" to explain user retention. The idea is simple: when a new model is the very first to solve a critical, previously unmet need for a group of users, it achieves a "perfect fit." These early users form "foundational cohorts" that stick with the model for the long term, showing far higher retention than users who adopt it later. Their workflows, tools, and habits become locked into the model that first solved their problem, making them highly resistant to switching to newer, even slightly better, alternatives. A clear example can be seen in the June 2025 cohort for Gemini 2.5 Pro and the May 2025 cohort for Claude 4 Sonnet, which retained around 40% of their users five months later; a far higher rate than subsequent cohorts. This contrasts sharply with models like Gemini 2.0 Flash, which, according to the report, failed to establish any foundational cohort, demonstrating that launching into a "good enough" market without a breakthrough capability results in high churn across all users. This phenomenon means that in the fast-moving world of AI, the "first-to-solve" advantage is incredibly powerful and durable. The true signal of a groundbreaking model isn't just hype or benchmarks, but the quiet formation of these sticky, foundational user groups who have found their perfect fit. Conclusion: The Real Story of AI Is Still Being Written The real-world usage of AI is far more nuanced, diverse, and surprising than the mainstream narrative suggests. The data shows an ecosystem driven as much by playful creativity as by productivity, one that is rapidly shifting toward complex, agentic workflows. It is a world where open source is a powerful force and where durable user loyalty is forged not by hype, but by being the first to solve a difficult problem. As AI moves from being a simple tool to a complex collaborator, what unexpected behaviors and killer apps will emerge next? Apple Podcast: HERE Spotify: HERE Apple Podcast: HERE Apple Podcast: Apple Podcast: HERE HERE Spotify: HERE Spotify: Spotify: HERE HERE
