Overview
-
Open vs. Closed Source Models: Open-weight models have grown to approximately 30% of total token usage by late 2025, up from negligible levels in late 2024. Chinese-developed models like DeepSeek and Qwen drove significant growth, reaching nearly 30% weekly share at peaks. The ecosystem has shifted from DeepSeek's near-monopoly to a pluralistic mix where no single open model exceeds 25% of OSS tokens.
- Major releases (DeepSeek V3, Kimi K2, GPT OSS) triggered sustained adoption spikes
- Medium-sized models (15-70B parameters) emerged as a growing category, balancing capability and efficiency
-
Rise of Agentic Inference: Reasoning-optimized models now represent over 50% of all token usage, up from negligible levels in early 2025. Tool-calling behavior has increased steadily, with models like Claude Sonnet and Gemini Flash leading tool-enabled workflows.
- Average prompt tokens per request quadrupled from ~1.5K to over 6K tokens
- Programming workloads drive the longest sequences, averaging 3-4x general prompt lengths
-
Category Distribution: Roleplay and programming dominate LLM usage across all models. Roleplay accounts for ~52% of open-source token usage, while programming has grown from 11% to over 50% of total volume through 2025.
- Anthropic's Claude dominates programming tasks with 60%+ market share
- Chinese OSS models are increasingly used for technical workloads, not just creative tasks
-
Geographic Patterns: Over 50% of OpenRouter usage originates outside the United States. Regional preferences vary significantly, with different model families gaining traction in different markets based on local offerings and language support.
-
Retention Analysis: The study identifies "foundational cohorts"—early users whose engagement persists far longer than later cohorts, termed the "Glass Slipper" effect. Early alignment between user needs and model characteristics creates lasting engagement.
Takeaways
OpenRouter and a16z analyzed over 100 trillion tokens of real-world LLM usage. Creative roleplay and coding assistance dominate usage, while open-source models have captured nearly one-third of the market.
The way developers and end-users engage with LLMs in the wild is complex and multifaceted.