Picking the right model for every request
No single model is best at everything, and no provider is up all the time. Router decides, request by request, which model should answer, and says why.

Most teams start with one model behind one API key. Then a cheaper model turns out to be good enough for half the traffic, a provider has a bad afternoon, a customer asks where their data is processed, and the one-key setup becomes a tangle of special cases in application code.
Router moves that decision out of the application. Every customer gets a dedicated Router instance, in a region they choose, that speaks the APIs their code already uses. Behind it sit their own provider accounts. In front of it, nothing changes but a base URL.
Drop in, don’t rewrite
Router accepts OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini generateContent and OpenAI Embeddings. The OpenAI and Anthropic SDKs, the Google Gen AI SDK and Claude Code all point at Router by changing their base URL and key. When a request goes to a provider that speaks the same format, it passes straight through; when it doesn’t, Router translates it, tool calls, images, structured output and reasoning included.
The providers are yours. Router calls OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex AI, the Gemini API, Cohere, Mistral and many OpenAI-compatible vendors and self-hosted servers with your own credentials, kept encrypted inside your instance. You pay providers directly, and nobody marks up your tokens.
How a route decides
A request names a model or a route. A route is a short pipeline that runs in memory and decides in under a millisecond. First it filters. Deployments the key isn’t allowed to use, that are over a price cap, that can’t handle what the request needs, such as tools, images or a long context, or that sit outside the residency policy are removed.
Then it classifies the request: what kind of task, how large, how hard. Rules match on that and choose a pool of models and an objective, such as the best quality, the lowest cost or the lowest latency. Router ranks what is left against the objective, using the latency it measures for each deployment as it goes, and keeps the rest of the ranking as the order to fall back in.
Names can be remapped too. A team that wants every small, quick request from a coding agent to land on a cheaper model can map those model names to a fast, inexpensive route without touching the agent.
When a provider fails
Providers rate-limit, slow down and go down. Router watches for all three. A deployment that keeps failing is moved out of first place until it recovers. If a call fails before the first byte comes back, Router tries the next candidate in the ranking, so the client never sees the error. Providers’ own rate-limit headers tell Router when a quota is spent, and it routes around that deployment until the quota resets.
Where your data goes
Every deployment has a processing location: a region, a geography, global, or unknown. A residency policy on the instance, a project or a key says which are allowed, and a project or key can only narrow the instance’s. Unknown, global and cross-region endpoints are refused unless you allow them.
The policy is checked when the configuration is saved, so nothing configured can break it, and again on every request, including failovers and fallbacks. A request with nowhere allowed to go is refused rather than sent somewhere it shouldn’t be. The instance’s own data, its configuration, credentials and usage, stays in the region where the instance runs.
Every answer explains itself
Each response carries headers naming the model and deployment that answered, the route and rule that chose it, the objective, the reason and how many attempts it took. When someone asks why a request went where it did, the answer is in the response itself, not on a dashboard you have to go and find.
One model will rarely be the right answer for everything you ask. With a router that chooses per request, and shows its work, you don’t have to pretend it is.
Written by Hexel Studio. Questions about anything here? Talk to us.


