Developer cuts AI agent costs from $87 to single digits weekly with custom token router

A developer running an OpenClaw AI agent was spending $87 per week after every tool call — including trivial cron job checks — was routed to an expensive paid model. Inspired by seeing OmniRoute report one billion tokens routed at zero cost, the developer built a pre-call classification system that assigns each request a call type and directs it to the most appropriate model tier. Simple tasks like heartbeat checks are handled by a local 9B model, while complex reasoning and code generation are reserved for paid cloud models. The routing logic also addressed a reliability issue where free OpenRouter models returned truncated output instead of standard 429 errors, which had silently caused three cron job failures in May. With deterministic routing applied before any model call, weekly costs dropped to single digits and the cron failures were eliminated.