OpenRouter's new provider router rewards quality per actual request cost

OpenRouter is a service that lets developers send requests to many AI models through one API. The same model is often offered by several hosting companies, which OpenRouter calls providers, at different prices and speeds, so for each request OpenRouter has to pick one of them. That choice is what this post is about. OpenRouter says it has shipped the biggest change to its provider router since the company was founded: the router used to prioritize "mostly based on price," and now it "rewards the best quality inference for the actual cost of a request."

OpenRouter's provider routing documentation, as it currently reads, spells out what that cost comparison involves. For each model in a request, the router compares what a good response would cost at each provider rather than list price alone, using measurements from recent production traffic: the price a provider actually charges once its real cache hit rate and mix of input and output tokens are counted, its output speed, the time until a response starts, how often it turns requests away for lack of capacity, how often requests fail, and whether its tool calls come back well formed. Providers close to the best share most of the traffic, with each request sent to one of them at random in proportion to those shares, while providers far behind keep a small share so the router notices if they improve. In the documentation's words, "a faster provider can charge more and still win your request." Those rules, as currently documented, match the post's claim that providers that deliver more value automatically win more traffic.

This is a different router from OpenRouter's Auto Router, which picks the model itself based on what OpenRouter users spend on for each kind of task. Provider routing works on a model that has already been named and only decides who serves it. Developers who would rather have the cheapest provider can still ask for it: the documentation says default routing is bypassed when a request sorts providers by price, throughput or latency, uses the :floor or :nitro shortcuts, sets an explicit provider order, or uses the developer's own provider key. The post's text gives no measurements of how much faster or more reliable requests have become.