OpenRouter’s 100 trillion token study reveals open-weight AI models are eating the market

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.

Open-weight doesn’t mean fully open-source in the traditional sense. The training data and methodology often remain proprietary. But the weights themselves are freely accessible.

What people actually use AI for

Creative roleplay and coding assistance emerged as the two dominant use cases across the platform. Roleplay alone represented more than 50% of open-weight model usage.

The study also flagged the rise of agentic workflows, where AI models aren’t just responding to single prompts but are chained together to complete multi-step tasks autonomously. This pattern has been growing across both open and closed model ecosystems.

Why this matters beyond the AI bubble

If open-weight models can capture a third of usage on a major routing platform, the moat around proprietary model providers is thinner than their valuations might suggest.

Chinese AI labs gaining nearly 30% weekly share through open-weight distribution is exactly the kind of development that will fuel ongoing debates about technology competition between the US and China. Open-weight distribution is, by design, harder to restrict through export controls or sanctions.

The a16z collaboration on this paper isn’t incidental. The venture firm has been one of the most aggressive investors across both crypto and AI, and its interest in mapping open-weight adoption patterns suggests it sees strategic opportunity at the intersection.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.