OpenRouter Data: Key AI Adoption Trends & Model Shifts

As AI systems move beyond text generation, understanding their real-world application and evolution becomes critical. OpenRouter has released a comprehensive report based on 13 months of data, analyzing over 100 trillion tokens to provide insights into the current state of artificial intelligence adoption and development.

Key Points

The analysis highlights several significant shifts in AI model usage and market dynamics:

Inference Model Dominance: Following the release of OpenAI's o1 inference model, there has been a notable shift from single-shot generation to multi-step deliberation and inference. Models optimized for inference now account for over 50% of token usage, establishing themselves as the default approach.
Closed-Source vs. Open-Source: While closed-source models continue to dominate, open-source alternatives are experiencing consistent growth. Projections indicate that by the end of 2025, open-source models will constitute approximately one-third of total usage. Chinese open-source models, in particular, have seen substantial growth, climbing from a 1.2% weekly share at the end of 2024 to nearly 30% in some weeks, with an annual weekly average of approximately 13.0%.
Diversification in Open-Source: The open-source landscape is transitioning from concentrated usage to a more diversified competitive environment. By the end of 2025, no single open-source model is expected to consistently exceed 25% of the market share, with 5–7 models distributed more evenly.
Rise of "Medium Models": Models categorized by parameter count (less than 15B as small, 15–70B as medium, greater than 70B as large) show a clear trend. Usage is shifting away from "small" models towards "medium" and "large" models, with "medium" models emerging as a new mainstay for "model-market fit."
Programming as a Growth Driver: Programming has become the fastest-growing category for AI application. Starting at approximately 11% at the beginning of 2025, it now exceeds 50%. For open-source models, usage is primarily split between role-playing (approximately 52%) and programming. In Chinese open-source models, role-playing accounts for about 33%, while programming and technical applications combined represent approximately 39%.
Increased Prompt and Sequence Lengths: The percentage of successful tool call requests increased throughout the year. The average prompt length grew from approximately 1.5K to over 6K (a fourfold increase), and completion length increased from approximately 150 to nearly 400 (a threefold increase). Average sequence length expanded from under 2,000 at the end of 2023 to over 5,400 by the end of 2025. Programming input length is a significant factor, being 3–4 times the overall average and serving as a primary driver for these increases.

Context

The report also details geographical and linguistic usage patterns. North America leads in continental share with 47.22%, followed by Asia at 28.61% and Europe at 21.32%. Asia's share significantly increased from approximately 13% to 31%. The top 10 countries by usage include the United States (47.17%), Singapore (9.21%), Germany (7.51%), and China (6.01%). English remains the dominant language, accounting for 82.87% of usage, with Simplified Chinese at 4.95% and Russian at 2.47%.

From a structural standpoint, the analysis indicates low price elasticity in the market, suggesting that it is not yet commoditized. Quality, reliability, and integration are noted as significantly outweighing price considerations. In practice, signs of "Jevons Paradox" are present, where efficient, low-cost models lead to longer contexts and more iterations, ultimately driving up total consumption.