Real-World AI Uses: Surprising Truths Revealed by Billions of Interactions

“`html

For the past year, we’ve been told that artificial intelligence is revolutionizing productivity by helping draft emails, generate code, or summarize documents. But is that really what’s happening? How are users actually leveraging AI chatbots like ChatGPT or Claude in their daily lives?

A landmark study conducted by OpenRouter has just shed light on real-world AI usage by analyzing over 100 trillion tokens, or billions of conversations with language models. The findings overturn many preconceptions about this technological revolution. What are the dominant use cases? How are trends evolving geographically? And most importantly, why do users keep returning to certain models over others?

An unprecedented study on LLM usage

OpenRouter is a multi-model inference platform that routes requests to over 300 models from 60 different providers, from OpenAI and Anthropic to open-source alternatives like DeepSeek and Meta’s LLaMA. With over 50% of its usage coming from outside the United States and serving millions of developers worldwide, the platform offers a unique cross-sectional view of AI usage.

The study analyzed the metadata of billions of interactions without accessing the textual content of conversations, thus preserving user privacy while revealing significant behavioral patterns.

Open-source AI models now represent approximately one-third of total usage by late 2025, with notable peaks following major releases.

Roleplay dominates real-world open-source AI usage

The most surprising discovery of this study concerns open-source models: over half of their usage is not dedicated to productivity, but to roleplay and narrative creation.

Yes, you read that right. While leaders of technology companies tout AI’s potential to transform the business world, users spend the majority of their time engaging in conversations with fictional characters, interactive fiction, and game scenarios. More than 50% of interactions with open-source models fall into this category, even surpassing programming assistance.

The data shows that users treat AI models as true structured roleplay engines, with 60% of roleplay tokens dedicated to specific game scenarios and creative writing contexts. This massive and largely invisible usage is redefining how AI companies design their products.

Programming experiences explosive growth

While roleplay dominates open-source, programming has become the fastest-growing category across all models. At the beginning of 2025, code-related requests represented only 11% of total usage. By year-end, that figure skyrocketed to exceed 50%.

This growth reflects the increasing integration of AI in software development. The average length of prompts for programming tasks has quadrupled, from approximately 1,500 tokens to over 6,000, with some code-related requests exceeding 20,000 tokens—equivalent to an entire codebase submitted for AI model analysis.

Programming-related requests have exploded, growing from 11% of total usage at the beginning of 2025 to over 50% by year-end.

Developers are no longer simply requesting simple code snippets. They are conducting sophisticated debugging sessions, architectural reviews, and multi-step problem solving. Anthropic’s Claude models dominate this space, capturing over 60% of programming-related usage through much of 2025, though competition is intensifying from Google, OpenAI, and open-source alternatives.

The spectacular rise of Chinese AI models

Another major revelation from the study: Chinese AI models now account for approximately 30% of global usage, nearly triple their 13% share at the beginning of 2025.

Models from DeepSeek, Qwen (Alibaba), and Moonshot AI have rapidly gained ground, with DeepSeek alone processing 14.37 billion tokens during the study period. This shift represents a fundamental change in the global AI landscape, where Western companies no longer hold undisputed dominance.

Simplified Chinese is now the second most common language for AI interactions globally with 5% of total usage, behind English at 83%. Asia’s share of AI spending has more than doubled, from 13% to 31%, with Singapore emerging as the second-largest country by usage after the United States.

The emergence of agentic AI

The study introduces a concept that will define the next phase of AI: agentic inference. This means that AI models are no longer just answering isolated questions. They execute multi-step tasks, call external tools, and reason over extended conversations.

The share of AI interactions classified as “optimized for reasoning” jumped from nearly zero at the beginning of 2025 to over 50% by year-end. This shift reflects a fundamental change: AI is transitioning from text generator to autonomous agent capable of planning and execution.

Instead of asking AI to “write a function,” users now ask it to “debug this codebase, identify the performance bottleneck, and implement a solution,” and it can actually do it.

The “glass slipper” effect and user loyalty

One of the study’s most fascinating discoveries concerns user retention. Researchers identified what they call the Cinderella “glass slipper” effect, a phenomenon where AI models that are “first to solve” a critical problem create lasting user loyalty.

When a newly released model perfectly matches an unmet need, early adopters stick around much longer than late adopters. For example, Google’s Gemini 2.5 Pro cohort from June 2025 retained approximately 40% of users in the fifth month, a substantially higher rate than later cohorts.

This finding challenges conventional wisdom about AI competition. Being first matters, but specifically being first to solve a high-value problem creates a sustainable competitive advantage. Users integrate these models into their workflows, making switching costly both technologically and behaviorally.

Price is not the deciding factor

Counterintuitively, the study reveals that AI usage is relatively price-inelastic. A 10% price decrease corresponds to only about 0.5 to 0.7% increase in usage.

Anthropic and OpenAI’s premium models charge between $2 and $35 per million tokens while maintaining high usage, while budget options like DeepSeek and Google’s Gemini Flash reach similar scale at under $0.40 per million tokens. Both coexist successfully.

The LLM market does not yet behave like a commodity. Users balance cost with reasoning quality, reliability, and breadth of capabilities. Quality and performance still command premium prices, at least for now.

What real-world AI usage reveals about the future

The OpenRouter study paints a portrait of real-world AI usage far more nuanced than industry narratives. Yes, AI is transforming programming and professional work. But it’s also creating entirely new categories of human-machine interaction through roleplay and creative applications.

The market is diversifying geographically, with China emerging as a major force. The technology is evolving from simple text generation to complex multi-step reasoning. And user loyalty depends less on being first to market than on being first to genuinely solve a problem.

Understanding these real usage patterns, not just benchmark scores or marketing claims, will be crucial as AI increasingly integrates into our daily lives. The gap between how we think AI is used and how it’s actually used is wider than most realize. This study helps close that gap.

Source: Artificial Intelligence News | OpenRouter State of AI Study

“`