If you want to know what people actually do with AI models, and not what benchmarks say they should do, OpenRouter just dropped the receipts. The company, in collaboration with Andreessen Horowitz (a16z), published a sprawling analysis of over 100 trillion tokens of anonymized LLM interaction data, covering roughly a year of real-world usage on its platform through late 2025.
The headline finding: open-weight models now account for approximately one-third of all token volume.
The rise of open-weight, and the China factor
The study, titled “State of AI: An Empirical 100 Trillion Token Study with OpenRouter,” was released on arXiv and represents one of the largest observational analyses of how people interact with large language models in the wild. Most prior research in this space has leaned toward qualitative surveys or synthetic benchmark comparisons. This one is built on raw usage data.
Chinese-developed open-weight models went from capturing roughly 1.2% of weekly token share in late 2024 to peaks near 30% at various points throughout 2025, averaging approximately 13% weekly share across that period.
