OpenRouter Users Sent 84% of Tokens to Non-Frontier Models in Aug. 10 Week

Developers using OpenRouter directed 84% of their token volume through non-state-of-the-art models during the week of Aug. 10, highlighting the role of lower-cost systems in workloads handled through the model-routing platform.

Six models accounted for about 80% of OpenRouter volume while delivering roughly 77% of frontier-model performance. The group averaged about $0.50 per million tokens, compared with $20 per million tokens for Claude Fable 5, or 2.5% of the cost.

“The six models users choose to generate the supermajority of those tokens deliver about 77% of the performance of the frontier,” venture capitalist Tomasz Tunguz wrote. “They cost 2.5% of what Claude Fable 5 does.”

The data suggests developers using OpenRouter may be balancing price and performance instead of selecting the highest-ranked model for every request. Many lower-cost models may be sufficient for routine workloads when their performance meets the task's requirements.

Frontier models can cost 10 to 50 times more per token than capable open-weight models. Open-weight models can be downloaded and modified, giving developers more control over deployment and potentially reducing operating costs.

OpenRouter functions as an API routing layer that provides access to models from multiple providers through one interface. Its standard and business plans provide access to more than 500 models and more than 80 providers.

Routing can account for cost, speed, availability and policy requirements. OpenRouter’s :floor option prioritizes the lowest-price provider for a selected model. Model inference pricing is passed through from providers without a markup, while the platform charges a 5.5% fee when users purchase credits.

The analysis covers traffic sent through OpenRouter and excludes requests made directly through OpenAI, Anthropic and Google services. It therefore measures model selection among router users rather than the entire AI market.

The pattern does not establish that frontier models are losing all demand or revenue. It shows that lower-cost models handled a large share of OpenRouter workloads during the measured week when their performance was sufficient.