What Are We Really Doing With AI? 5 Surprising Truths From 100 Trillion Token Analysis
In the popular imagination, the AI revolution boils down to a few key applications: writing emails, generating code, or summarizing long documents. It’s a productivity narrative where AI is our digital assistant, helping us work faster and more efficiently.
However, an analysis of an unprecedented scale of 100 trillion tokens from the "State of AI" report, conducted by a16z (Andreessen Horowitz) and OpenRouter Inc., reveals a completely different, far more surprising picture. The reality of how the global community actually uses language models is more complex, creative, and fundamentally human than we might have guessed.
This article will take you behind the scenes and, correcting a few popular myths, reveal five counterintuitive truths about how the world really uses AI. We’ll start with a discovery that turns our assumptions about the technology’s dominant uses upside down.
1. The Hidden King of AI: It’s Not Productivity, It’s Roleplay
The most surprising conclusion from the analysis is a fundamental truth: the dominant use case for open-source models isn't business tasks, but creative entertainment.
The analysis of the "Category Breakdown of OSS Models over Time" chart is unequivocal: the "Roleplay" category consistently accounts for over 50% of all token usage in open-source models. This means the community of developers and AI enthusiasts dedicates more computing power to creating interactive stories and simulations than to 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 roleplay in fantasy worlds.
This finding is significant because it unmasks a massive, grassroots consumer market that values creativity and freedom. Open-source models, often less constrained by rigid content filters, have become the perfect testing ground for communities creating interactive narratives, games, and character simulations.
2. The End of the Chatbot Era: AI Becomes an "Agent" for Special Tasks
We are observing a fundamental shift in the nature of interaction with AI. The median LLM query is no longer a simple question, but part of a structured, agentic loop. We are moving from simple Q&A to entrusting models with complex, multi-step tasks, reflecting a human desire to delegate more meaningful analytical and creative projects.
This new paradigm, known as agentic inference, treats AI not as a text generator, but as an autonomous agent capable of planning, using tools, and executing multi-step tasks to achieve a goal. The report provides three hard pieces of evidence for this:
- Dominance of Reasoning Models: The ability to plan is becoming the standard. The "Reasoning vs Non-Reasoning Token Trends Over Time" chart shows that models capable of reasoning, which were a rarity even in early 2025, now process over 50% of all token traffic.
- Explosion of Query Complexity: We are communicating with AI in a much more complex way. Since the beginning of 2024, the average prompt length has quadrupled. The main driver of this growth is programming tasks, where contexts regularly exceed 20,000 tokens, feeding models entire code snippets for analysis.
- Increasing Tool Usage: There is a growing use of "Tool Calling," which allows AI models to interact with external systems and APIs. This means AI isn't just generating text; it is actively performing actions on our behalf.
Artificial intelligence serves less as a passive content generator and more as an active partner in work, capable of independent planning and execution.
3. The "Glass Slipper" Effect: Why Early Adopters Are the Most Loyal
In the dynamic world of AI, user retention works differently than in traditional software. The report describes this phenomenon with an apt metaphor—the Cinderella 'Glass Slipper' effect.
The concept is simple: when a new model is the first to solve a critical, previously unsolvable problem for a specific group of users, it earns their lasting loyalty. This first group, dubbed the "founding cohort," builds entire workflows, data processes, and habits around the model, generating immense switching friction and leading to lasting vendor lock-in. The cost (in time and effort) of switching to a newer model becomes prohibitive.
The report illustrates this with concrete examples:
- Positive Example: Retention charts for Claude 4 Sonnet (May 2025 cohort) and Gemini 2.5 Pro (June 2025 cohort) show that early users of these models stayed with them significantly longer than later groups. They found their "perfect fit."
- Negative Example: In contrast, the Llama 4 Maverick model did not find its "founding cohort." No user group discovered a unique use case for it, resulting in uniformly low retention across the board.
Interestingly, in the case of DeepSeek models, a "Boomerang Effect" was observed: users who initially left returned after testing alternatives. This is proof that the model found a unique fit for specific, niche tasks that the competition did not offer.
4. Price (Mostly) Doesn't Matter: What Really Decides AI Choice
Intuition suggests that lower prices should lead to higher usage. However, AI market analysis shows that demand is largely price inelastic. The report states this plainly: a 10% price drop corresponds to a usage increase of only 0.5–0.7%.
The "Log Cost vs Log Usage by Category" chart proves that the market is divided into four distinct segments:
- Mass Volume Leaders: Categories like Roleplay and Programming are characterized by huge usage volume at low to medium cost. This is where efficient open-source models reign, offering the optimal balance of performance and cost-efficiency for price-sensitive workloads.
- Premium Workloads: Tasks in Technology and Science categories feature high costs and high popularity. Users are willing to pay a premium for top quality, reliability, and advanced reasoning capabilities.
- Specialized Experts: Niche, expensive applications like Finance, Health, and Marketing. In these fields, precision and accuracy are more important than cost, and usage volume is naturally lower.
- Niche Tools: Categories like Translation, Legal, or Trivia are commoditized, cheap services with lower volume, where "good enough" quality is available for a small price.
5. Changing of the Guard: AI's Center of Gravity Shifts East
Innovation in artificial intelligence is no longer the exclusive domain of Silicon Valley. The report provides strong evidence for the ongoing globalization and decentralization of the AI market.
- The Rise of Asia: Based on the "Spend Volumes by World Region Over Time" chart, we can see that Asia's share of global AI spending grew from 13% in the first weeks of the analysis (late 2024) to about 31% in the final period (late 2025). That’s more than a twofold increase.
- The Power of Chinese Open Source Models: The analysis shows that Chinese open-source models, such as DeepSeek and Qwen, have significantly contributed to the growth of the entire open ecosystem. They quickly gained significant market share, often competing in quality with their Western counterparts.
Global competition accelerates innovation, democratizes access to technology, and gives users more choice. The future of AI will depend on the models' ability to operate in different languages and understand diverse cultural contexts. The era of single-region dominance is coming to an end.
Conclusion: The True Face of the AI Revolution
The analysis of 100 trillion tokens shows that the actual use of artificial intelligence is far more human, creative, and diverse than the dominant tech narrative suggests. It turns out that alongside optimizing work, we have an equally strong desire to create, play, and explore new worlds.
AI is becoming not just a productivity tool, but a medium for our imagination. We are moving from simple commands to complex, agentic interactions, and model success is determined by the ability to durably solve real problems, not price alone.
If AI is already revealing our unexpected needs for creativity and interaction today, what other surprising aspects of humanity will it uncover in the future?
Aleksander
Sources:
About the Author
Dyrektor ds. Technologii w SecurHub.pl
Doktorant z zakresu neuronauki poznawczej. Psycholog i ekspert IT specjalizujący się w cyberbezpieczeństwie.
Powiązane artykuły
Dzieci w Sieci 2025: Patostreamy, "Sleepy Chicken" i Prawo, które Karze Ofiary.
Liczba incydentów bezpieczeństwa z udziałem dzieci przebiła sufit – ponad 600 tysięcy zgłoszeń. Analizujemy dane NASK i Policji: od śmiertelnych wyzwań na TikToku, przez wyłudzenia w Robloxie, aż po dramatyczne luki prawne w sekstingu.
Vibe Coding: Rewolucja czy Rosyjska Ruletka? Mroczna strona programowania z AI
Wszyscy „czują vibe", ale nikt nie czyta kodu. Analizujemy zjawisko Vibe Codingu, plagę Slopsquattingu i to, jak AI po cichu degraduje bezpieczeństwo naszych aplikacji.
Dlaczego Jesteśmy Drugim Najczęściej Atakowanym Krajem na Świecie?
Najnowszy raport ESET rzuca światło na alarmującą pozycję Polski w globalnym rankingu cyberzagrożeń. Jesteśmy na drugim miejscu na świecie pod względem liczby ataków. Analizujemy, co to oznacza i jak wygląda krajobraz zagrożeń, od ransomware po ataki na infrastrukturę krytyczną.
Komentarze
Ładowanie komentarzy...