What Do People Really Use AI For? 5 Surprising Truths from 100 Trillion Tokens

In the popular imagination, the AI revolution comes down to a few core use cases: writing emails, generating code and summarizing long documents. It's a productivity story, with AI as a digital assistant that helps us work faster and smarter.

But an analysis of an unprecedented 100 trillion tokens ([editor's note] the report covers data sent through the API and excludes "wrapped" tools offered directly by vendors) in the "State of AI" report by a16z (Andreessen Horowitz) and OpenRouter Inc. paints a very different and far more surprising picture. How the global community actually uses language models is more complex, more creative and more human than we might have guessed.

This article takes you behind the scenes, debunks a few popular myths and reveals five counterintuitive truths about how the world really uses AI. We start with the finding that most upends our assumptions about what this technology is mainly used for.

1. AI's Hidden King: It's Not Productivity, It's Roleplay

The most surprising finding in the analysis is this: the main use of open-source models isn't business work. It's creative entertainment.

The "Category Breakdown of OSS Models over Time" chart leaves no room for doubt: the "Roleplay" category (we wrote about its psychological side here) consistently accounts for more than 50% of all token usage on open-source models. In other words, the AI developer and enthusiast community spends more compute on interactive stories and simulations than on any other task.

It turns out that while Silicon Valley was building AI to optimize spreadsheets, the world was using it to write epic sagas and role-play in fantasy worlds.

This matters because it reveals a huge grassroots consumer market that values creativity and freedom. Open-source models, which often have looser content filters, have become the ideal testing ground for communities building interactive stories, games and character simulations.

2. The End of the Chatbot Era: AI Becomes an "Agent" for Specialized Tasks

We're seeing a fundamental change in how people interact with AI. The median LLM request is no longer a simple question. It's one step in a structured, agentic loop. We're moving from simple Q&A to handing models complex, multi-step tasks, which shows that people want to delegate more substantial analytical and creative work.

This new paradigm, known as agentic inference, treats AI as an autonomous agent rather than a text generator: one that can plan, use tools and carry out multi-step tasks to reach a goal. The report offers three pieces of hard evidence:

• Reasoning models dominate: The ability to plan is becoming standard. The "Reasoning vs Non-Reasoning Token Trends Over Time" chart shows that reasoning models, still rare in early 2025, now handle more than 50% of all token traffic.
• Prompts are getting much more complex: Our requests to AI are far more elaborate than before. Average prompt length has grown fourfold since early 2024. Programming tasks drive most of that growth, with contexts regularly exceeding 20,000 tokens as developers feed models whole sections of code to analyze.
• Tool use keeps growing: More and more requests use "Tool Calling," a feature that lets AI models interact with external systems and APIs. AI doesn't just generate text anymore. It takes action on our behalf.

AI is no longer a passive content generator. It's becoming an active work partner that can plan and carry out tasks on its own.

3. The "Glass Slipper" Effect: Why Early AI Adopters Are the Most Loyal

In the fast-moving AI market, user retention works differently than in traditional software. The report describes this with a fitting metaphor: the Cinderella "Glass Slipper" effect.

The idea is simple. When a new model is the first to solve a critical, previously unsolvable problem for a particular group of users, it wins their lasting loyalty. This first group, which the report calls the "foundational cohort," builds entire workflows, data pipelines and habits around the model. That creates heavy switching friction and lasting vendor lock-in, and the time and effort needed to move to a newer model become prohibitive.

The report illustrates this with concrete examples:

• A positive example: Retention curves for Claude 4 Sonnet (May 2025 cohort) and Gemini 2.5 Pro (June 2025 cohort) show that early users stayed with these models much longer than later groups did. They had found their "perfect fit."
• A negative example: Llama 4 Maverick, by contrast, never found its foundational cohort. No user group discovered a unique use for it, so retention was uniformly low across all cohorts.

Interestingly, DeepSeek models showed a "Boomerang Effect": users who had initially churned came back after trying the alternatives. That's evidence the model had found a unique fit for specific, niche tasks that competitors didn't offer.

4. Price (Mostly) Doesn't Matter: What Really Drives the Choice of an AI Model

Intuition says a lower price should mean more usage. But the AI market analysis shows that demand is largely price-inelastic. The report puts it plainly: a 10% price cut translates into only a 0.5–0.7% increase in usage.

The "Log Cost vs Log Usage by Category" chart shows that the market splits into four distinct segments:

1. High-volume mass-market leaders: Categories such as Roleplay and Programming combine huge usage volumes with low-to-medium costs. This is where efficient open-source models dominate, offering the best balance of performance and cost for price-sensitive workloads.
2. Premium workloads: Tasks in the Technology and Science categories are both expensive and popular. Users are willing to pay a premium for top quality, reliability and advanced reasoning.
3. Specialized experts: Niche, expensive use cases such as Finance, Health and Marketing. In these fields, precision and accuracy matter more than cost, and usage volumes are naturally lower.
4. Niche utilities: Categories such as Translation, Legal and Trivia are commoditized, low-cost, lower-volume services where "good enough" quality is available cheaply.

5. A Changing of the Guard: AI's Center of Gravity Is Shifting East

AI innovation is no longer the exclusive domain of Silicon Valley. The report provides strong evidence that the AI market is becoming more global and decentralized.

• Asia's rise: The "Spend Volumes by World Region Over Time" chart shows Asia's share of global AI spending growing from 13% in the first weeks of the analysis (late 2024) to about 31% in the most recent period (late 2025). That's more than double.
• The strength of Chinese open-source models: The analysis shows that Chinese open-source models such as DeepSeek and Qwen drove much of the growth of the whole open-source ecosystem. They quickly won significant market share and often match their Western counterparts on quality.

Global competition speeds up innovation, broadens access to the technology and gives users more choice. The future of AI will depend on how well models work across languages and understand different cultural contexts. The era of one region's dominance is coming to an end.

Conclusion: The True Face of the AI Revolution

The analysis of 100 trillion tokens shows that real-world AI use is far more human, creative and diverse than the dominant tech narrative suggests. Alongside making our work more efficient, we want just as much to create, to play and to explore new worlds.

AI is becoming more than a productivity tool. It's also a medium for our imagination. We're moving from simple commands to complex, agentic interactions, and a model's success depends on its ability to keep solving real problems, not on price alone.

If AI is already revealing needs for creativity and interaction we didn't expect, what other surprising sides of human nature will it uncover next?

Aleksander

Sources:

• "State of AI" report - OpenRouter Inc.