Model Routing — Haiku for Speed, Sonnet for Reasoning, Opus for Architecture

Key Takeaways

• Treat model selection as a routing policy, not a one-time configuration — classify each call by latency class, reasoning depth, and context length before dispatching, exactly as an API gateway applies route rules based on path and method.
• Haiku is the edge-cache tier: sub-second, high-throughput, low reasoning complexity — entity extraction, intent classification, slot-filling; Sonnet is the application tier: balanced latency and multi-step reasoning for summarisation, RAG synthesis, and moderate-length document analysis; Opus is the batch-processing or async tier: complex architecture review, cross-document regulatory inference, multi-constraint planning where latency tolerance exists.
• Long-context scenarios demand model-capability validation before routing — a 100K-token contract document routed to Haiku may process without error but silently degrade inference quality; always benchmark reasoning fidelity, not just throughput, during model selection for context-heavy workloads.
• Memory anchor — API gateway tiered routing: just as your gateway applies latency-sensitive fast-path vs. enriched-path routing rules, your model router applies Haiku-fast-path, Sonnet-standard-path, Opus-deep-path based on request classification at dispatch time.