How ModelPlane's Five-Stage Routing Engine Handles Every LLM Request

ModelPlane is an LLM gateway that routes every AI request through five sequential stages: authentication, config resolution, a billing gate, routing, and asynchronous accounting. The gateway resolves a user-facing Bearer token into a tenant context and strips it before forwarding the request upstream, ensuring provider credentials are never exposed to clients. Instead of hard-coded provider model IDs, users define named 'model groups' that map to multiple backends and routing strategies, giving teams provider-agnostic control over their AI calls. A pre-request credit check blocks requests with insufficient balance and returns a 402 error, while usage accounting is handled asynchronously after tokens are returned to keep the hot path fast. The design aims to prevent common gateway pitfalls such as credential leakage, dropped usage data, and latency introduced by synchronous billing operations.