State of AI: An Empirical 100 Trillion Token Study with OpenRouter

The Evolution of LLMs

• 💡 The field has shifted from single-pass pattern generation to multi-step deliberation inference since the release of the o1 model on December 5th, 2024.
• 🧠 This means models can now stop, think, sketch out a plan, and check their own work, acting as analytical partners rather than just prediction engines.
• 📊 An empirical study by OpenRouter analyzed over 100 trillion tokens of real-world LLM interactions, spanning a year up to November 2025.

Key Usage Trends

• 🎭 Creative roleplay accounts for over 50% of open-source model token usage, highlighting a preference for creative freedom over commercial safety layers.
• 💻 Programming is the second largest use case and the dominant driver of complexity, with queries often exceeding 20,000 input tokens for tasks like feeding entire codebases.
• 🚀 The "medium market" of models between 15 and 70 billion parameters (e.g., Quen 2.5, Coder 32B) is finding optimal "model market fit" for balancing capability and efficiency.

Rise of Agentic Inference

• ⚡ Agentic inference is becoming the default, enabling models to plan, reason, and execute complex multi-step workflows.
• 📈 This is evidenced by over 50% of tokens routed through reasoning-optimized models and consistent adoption of tool use (e.g., search engines, code interpreters).
• 📊 The average number of prompt tokens per request has grown fourfold, indicating users are providing vast amounts of context like code and logs.

Market Dynamics & Retention

• 💎 The Cinderella "Glass Slipper" effect describes persistent user retention when a specific high-value workload perfectly matches a model's capabilities.
• 💰 There is a very weak correlation between price and demand for LLMs; users prioritize a model's ability to solve problems over its cost.
• 🌍 The market is becoming increasingly global, with Asia's share of global spend more than doubling and Simplified Chinese becoming the largest non-English language for tokens.

Future Outlook

• 🎯 The industry is moving away from optimizing for cost per token towards optimizing for the cost per successful outcome.
• ✅ The market is highly segmented into archetypes like mass market volume drivers, efficient giants, premium leaders, and premium workload categories.
• 🤝 A multi-model, pluralistic ecosystem is emerging, where integration and unique capabilities trump cost, forcing providers to differentiate.