2025 AI Ecosystem: OpenRouter Report Reveals Roleplay Dominance, Open-Weight Surge, and Agentic Shift

The year 2025 marked a pivotal period for artificial intelligence, characterized by the widespread adoption of reasoning models and a significant restructuring of the LLM ecosystem. Catalyzed by OpenAI’s ‘01’ in late 2024 and Deepseek’s ‘R1’ in early 2025, the industry witnessed a rapid shift from single-pass pattern generation to multi-step deliberation. A comprehensive empirical study by OpenRouter, leveraging over 100 trillion tokens of real-world LLM inference, provides unique insights into these trends, revealing that open-weight models have grown to command approximately one-third of total token usage by late 2025. This surge was notably driven by releases like Deepseek V3, Kimi K2, and GPT-OSS, with Chinese open-weight models experiencing explosive growth from 1.2% to nearly 30% in some weeks.

Surprisingly, the study identifies creative role-playing as the dominant use case for open-weight models on OpenRouter, accounting for over 50% of tokens, closely followed by programming. The average number of prompt tokens per request quadrupled, with completion tokens tripling, largely driven by increasingly complex programming tasks. Agentic inference has become the new default, with over half of all tokens routed through reasoning-optimized models, signifying a shift towards structured, tool-invoking interactions. While proprietary models still hold the majority share, the open-weight market has diversified from Deepseek’s initial near-monopoly, with medium-sized models (15-70B parameters) gaining significant traction. OpenRouter’s data also indicates that model quality and capabilities, rather than just cost, are the primary drivers of adoption, challenging conventional assumptions about market dynamics in the rapidly evolving LLM landscape.