Routing every LLM request to the same frontier model wastes money on simple tasks. DigitalOcean's Inference Router uses a purpose-built Mixture-of-Experts model called Plano-Orchestrator (30B total, ~3B active parameters) to classify incoming requests and dispatch them to the most cost-effective model in ~200ms. Benchmarks show it achieves 87.84% routing accuracy across multi-turn conversations, outperforming GPT-5.1 and Claude Sonnet 4.5 on the same task. A worked example shows a 59% reduction in output token costs for a typical coding agent workload. The article covers the MoE architecture rationale, the X-Model-Affinity header for session pinning to preserve KV cache, task description best practices, latency tradeoffs, and how to integrate with one string change in an OpenAI-compatible API call.
