One year ago, the first widely adopted reasoning model, OpenAI’s o1, launched and kicked off a new era of LLMs centered on reasoning, planning, and tool use. In that short time, single-pass, autoregressive systems that once defined AI deployment have started to give way to real-world usage of multi-step reasoning models and agentic workflows. Meanwhile, progress has not been limited to proprietary frontier labs; the number, quality, and adoption of open-weight models has accelerated, and they now serve as serious building blocks for production applications.

Until now there has not been a great way to understand key LLM trends rigorously at scale. We are excited to be releasing, in partnership with a16z, the 2025 State of AI Report. This is the largest empirical look yet at how developers and organizations actually use language models in the real world.

OpenRouter has a bird’s-eye view of global LLM traffic, and in collaboration with a16z, we analyzed metadata from more than 100 trillion tokens across our multi-model platform in the past year. We hope the insights from this study will help model builders, business leaders, application developers, and researchers better understand how AI is used in practice today, and make better decisions about where to invest, what to build, and where AI adoption is heading next.