State of AI: An Empirical 100T Token Study with OpenRouter
6 days ago
- #LLM Usage
- #Agentic Inference
- #OpenRouter
- The study analyzes over 100 trillion tokens of real-world LLM interactions using the OpenRouter platform.
- Open-source models (OSS) have grown to account for about one-third of total usage by late 2025, with Chinese-developed models showing significant adoption.
- Creative roleplay and programming assistance dominate open-source model usage, accounting for more than half of all OSS token volume.
- The study identifies a shift towards agentic inference, where models perform multi-step reasoning, tool integration, and iterative refinement.
- Programming has become the most expanding category, with models increasingly used for code generation, debugging, and technical reasoning.
- Retention analysis reveals foundational cohorts of users who persist longer due to workload-model fit, termed the 'Cinderella Glass Slipper' effect.
- Global usage shows North America as the largest region, but Asia's share has more than doubled, reaching nearly 31% by late 2025.
- Cost vs. usage analysis indicates that proprietary models dominate high-value tasks, while open models excel in cost-sensitive, high-volume applications.
- The study highlights the importance of empirical usage data to inform model development, deployment strategies, and infrastructure planning.