Korva is a scroll of the internet’s best ideas.
OpenRouter's rankings skew to open models
OpenRouter earns a fee on top of every token. That makes it useful mainly to people running open-source models, who want failover and the other value-added services it provides. Almost nobody routes proprietary models through it, so its numbers cover a narrow slice of usage.
Most inference spend still goes through first-party APIs to proprietary models. Anthropic's revenue has been a hockey stick, and Chinese open-source models are still a small fraction of the total inference spend in the world.
The ranking reflects who needs a layer between themselves and the model. Open-source users do, and users of closed models buy from the maker directly, so the closed-source business is far from being cannibalized yet.
How does a model router like OpenRouter work?
A router offers one API in front of many model providers and forwards each request to whichever provider hosts the model asked for. If a provider goes down, the router can retry the request on another provider. A single account and bill covers all of them. The request format modeled on OpenAI's chat completions API became a de facto standard that most providers copy.
Why do enterprises call model APIs directly?
Large customers usually sign contracts with the model maker for volume discounts, data-processing terms, and support. Many also buy through the cloud platforms they already use, such as Amazon Bedrock, Google Vertex AI, and Microsoft Azure, which offer models under existing billing and security arrangements. Those channels never pass through a third-party router. As a result, a router's usage data does not capture them.
Why can usage rankings mislead?
A ranking built from one platform's traffic reflects only the people who use that platform. Selection effects of this kind are a standard problem in statistics, because the sample is not drawn from everyone. Token counts also measure volume rather than value, since a model used for thousands of cheap calls can outrank one used for a few expensive ones. Public leaderboards that rank by user votes, such as Arena, measure a different thing from traffic-based ones.
More from 20VC with Harry Stebbings
