How much do open-weight models actually get used in practice?

An empirical study of over 100 trillion tokens of real-world usage on the OpenRouter inference platform, which routes requests across "more than 300+ active models from over 60 providers," reports that "even though OpenAI and Anthropic models lead in many programming and knowledge tasks, open source models like DeepSeek and Qwen collectively served a large portion of total tokens (sometimes over 30%)." The study frames this as evidence of "A Multi-Model Ecosystem": "no single model dominates all usage."

The mechanism the authors give is routing behavior, not model quality alone. Because OpenRouter sits as a neutral layer across closed APIs and open-weight deployments, it captures "a representative cross-section of how developers and end-users actually invoke language models for various tasks," including cases where a cheaper or differently-tuned open model is chosen even when a closed model performs better on a given task category. The authors argue the resulting heterogeneity "is not a transient phase but a fundamental property of the market," because users "select systems along multiple axes such as capability, latency, price, and trust depending on context." Practically, they read this as meaning developers should keep "flexibility, integrating multiple models and choosing the best for each job," and that incumbent providers face competition "from unexpected places."

This gives demand-side, platform-scale confirmation to the supply-side case made in Can small language models handle most agent tasks?, which argues heterogeneous, cost-driven model selection is the economically rational architecture for agentic systems. Where that paper reasons from inference-cost structure to a predicted heterogeneous future, the OpenRouter data shows a heterogeneous market already in place today, with open models capturing a double-digit share of total tokens without displacing closed models from the task categories where they lead. The companion finding in Does LLM price alone determine which model users choose? supplies part of the explanation for why the split persists rather than collapsing toward the cheapest option.

The study does not establish that this 30%-plus figure generalizes beyond OpenRouter's own user base: its authors note the dataset reflects "a single platform" over "a finite time window," excludes "enterprise usage, locally hosted deployments, or closed internal systems," and derives task categories from a classifier run on only "approximately 0.25% of all prompts and responses." The figure is therefore a sampled, platform-specific estimate of token share, not a census of global model choice, and it says nothing about the revenue, margin, or total compute cost behind those tokens — only about which models users routed requests to.