What 100 Trillion Tokens Reveal About How People Actually Use AI

There's a lot of speculation about how people use AI. Productivity boosters. Coding assistants. Research tools. The assumptions tend to skew toward efficiency and output—getting more done in less time.

A new research paper from A16Z and OpenRouter offers something more useful than speculation: data. The study analyzed over 100 trillion tokens of real-world AI usage across the past year. To put that in perspective, a token is roughly a word or word piece, so we're talking about billions of conversations and prompts. This is one of the largest empirical studies of how people actually interact with AI tools.

The findings challenge some common assumptions. And for association leaders trying to understand where AI adoption is headed, the patterns here are worth paying attention to.

Role Play Dominates

Here's the finding that surprised a lot of people: over 50% of open source model usage is for role play and creative storytelling.

Not productivity. Not coding. Interactive fiction, character conversations, and entertainment.

Programming does come in second, so the productivity use cases are real. But the dominant use of these tools is people engaging with AI for creative and social experiences. They're building fictional worlds, having conversations with characters, exploring narratives.

What does this tell us? People are drawn to AI for reasons beyond efficiency. There's something about the interactive, responsive nature of these tools that people find genuinely engaging. They're not just using AI to save time—they're using it because they enjoy the experience.

For associations, this is worth sitting with. Member engagement strategies often focus on delivering information efficiently. And that matters. But the data suggests people respond to AI experiences that feel interactive and personalized, not just transactional. The question isn't only "how do we help members get answers faster?" It might also be "how do we create AI-powered experiences members actually want to return to?"

That's a different design challenge.

Open Source Is Surging

A year ago, open source AI models handled a negligible share of total usage. Today, they account for roughly 30% of all token volume.

Models like DeepSeek and Qwen have driven much of this growth. Chinese open source AI specifically went from 1.2% to nearly 30% of usage in some weeks. That's a dramatic shift in a short window.

The practical implication: the AI landscape is diversifying fast. No single company or model has a permanent lock on the market. The tools available today will be different from the tools available in six months, often with expanded capabilities at lower cost.

For associations evaluating AI investments, this reinforces the value of flexibility. Building your entire strategy around one provider's ecosystem carries risk. The organizations that will navigate this well are the ones building internal knowledge and adaptability rather than betting everything on a single platform.

It also means the barriers to AI adoption keep dropping. Capabilities that required expensive enterprise contracts two years ago are increasingly available through open source alternatives. Your smaller peer organizations aren't as far behind as you might think.

The Rise of Agentic Inference

This finding has significant implications for what AI can actually do.

Usage is shifting from single-question, single-answer interactions to multi-step reasoning workflows. AI systems that don't just respond, but plan, use tools, validate their work, and iterate. The study found that reasoning models now represent half of all usage. And prompt length has quadrupled—from about 1,500 tokens to over 6,000 tokens on average.

What's happening beneath the surface is even more dramatic. A single user inquiry might trigger dozens of separate AI prompts behind the scenes. One email asking a simple question could result in 300,000 to 700,000 tokens of processing as the system researches the person's history, checks relevant databases, validates facts, and refines the response.

Think about what that means practically. Imagine a member emails your association asking "what time does the meeting start?" A vague question. A traditional approach requires a staff member to research who this person is, what events they're registered for, and make an educated guess about which meeting they mean.

An agentic AI system can do all of that automatically. Pull the member's registration history. Check recent communications for context. Search your knowledge base. Cross-reference upcoming events. Then craft a response that not only answers the question but anticipates related needs—directions, parking, session recommendations.

That's not a chatbot answering FAQs. That's an AI agent doing research and reasoning. The ceiling on what these systems can handle is rising fast, and the shift from "assistant" to "agent" is well underway.

The Glass Slipper Effect

The study surfaced an interesting pattern in user retention. When a new model launches, there's a narrow window where it can capture users whose specific workflow finally fits. Once that fit happens, those users become deeply loyal and stick around through subsequent releases and updates.

Early adopters who find the right match have dramatically higher retention than people who try a tool later. The researchers called this the "glass slipper effect"—the moment when a user finds the AI that fits their particular needs.

There's a strategic lesson here, both for choosing AI tools and for positioning AI offerings to your members.

On the tool selection side: if you've been waiting for AI to mature before investing time in adoption, recognize that early fit creates stickiness. The organizations experimenting now are building familiarity and workflows that will compound over time. Waiting for the "right" moment means potentially missing the window where your team develops real fluency.

On the member side: associations have an opportunity to be the resource that helps people in their industry find their glass slipper. If you can guide members to AI tools and approaches that fit their specific professional context, you become part of their adoption story. That's a relationship with staying power.

But the window matters. If members figure out AI on their own—or worse, get guidance from a competitor—that loyalty forms elsewhere. The first-mover advantage here isn't about being the fastest to adopt AI internally. It's about being the trusted guide for your members as they navigate this shift.

Other Patterns Worth Noting

A few additional findings round out the picture.

Anthropic's Claude handles over 60% of programming-related AI usage. That's the most concentrated category by any single provider. If your members work in technical fields, that's relevant context for what tools they're likely already using.

Geography is shifting. North America now accounts for less than half of global AI usage. Asia has grown from 13% to 31%, with significant increases in Singapore, Germany, China, and South Korea. AI adoption is a global phenomenon, and the center of gravity is moving. If your association has international membership, their AI context may look quite different from the North American narrative.

What This Means for Associations

The gap between how we talk about AI and how people actually use it is significant.

The headlines focus on productivity and automation. The data shows creativity and engagement leading the way. The assumption is that AI usage is concentrated among technical professionals. The reality is that open source tools have democratized access and usage is spreading globally.

Most importantly, the shift toward agentic, multi-step AI means the ceiling on what's possible keeps rising. The AI tools of 18 months ago couldn't do what today's systems can do. And 18 months from now, the landscape will shift again.

For association leaders, a few things follow from this:

Understanding actual usage patterns matters. If you're designing AI-powered member experiences based on assumptions rather than data, you might be solving for the wrong things. People want engagement and interaction, not just efficiency.

Flexibility beats commitment to a single platform. The market is moving too fast and diversifying too quickly to lock in. Build internal capabilities and adaptability.

The agentic shift changes what's possible. Problems that seemed unsolvable due to scale—personalized member service, individualized guidance, responsive support—are increasingly within reach. The constraint was never willingness. It was the resource limitation of human attention. That constraint is loosening.

Your role as a guide matters more than your internal adoption. Yes, your organization should be exploring AI. But the bigger opportunity is becoming the trusted resource for your members as they figure this out for their own careers and organizations. That's the relationship that creates lasting value.

The data from 100 trillion tokens tells a story about where AI is actually going. It's more creative, more distributed, more capable, and more accessible than the headlines suggest. Associations that pay attention to the real patterns—not the assumed ones—will be better positioned to serve their members through what comes next.