GPT-4o-mini to o3-mini: 3-Tier Complexity Router

Engineering analyses warn that deploying o1 or o3 blindly across all application traffic is an anti-pattern that can multiply API costs by 10x to 20x and increase P95 latency from 1.2s to 35+s. A production routing pattern uses gpt-4o-mini (which achieves first-token latency under 500ms at typical load) to handle ~80% of simple tasks, routes standard multimodal/drafting tasks to gpt-4o, and dispatches hard reasoning tasks to o3-mini (which is estimated at 14x cheaper than o1 with comparable benchmark performance).

Router Prompt (gpt-4o-mini, temp=0): Classify the incoming request complexity into one route:
- SIMPLE (80% tier: classification, routing, summarization, entity extraction) -> Route to gpt-4o-mini
- STANDARD (coding, analysis, drafting, multimodal <2s latency) -> Route to gpt-4o
- HARD (math, scientific/financial/legal multi-step reasoning) -> Route to o3-mini