Why the 60 / 30 / 10 split works

Most engineering teams default to routing all traffic to the newest frontier model. This is financially unsustainable at scale.

When a new model ships, the default behavior is to update the API endpoint everywhere. It's easier than evaluating cheaper alternatives. But as usage scales, this "lazy routing" becomes a massive cost center.

You don't need frontier reasoning to extract a date from a receipt, classify a ticket, or summarize a chat log. You need a fast, cheap model that executes the instruction reliably.

The benchmark routing strategy

60%: Small models

Llama 3 8B, Mistral, GPT-4o-mini (basic tasks). Extraction, classification, structured output, routine chat.

30%: Medium models

Claude 3.5 Haiku, GPT-4o-mini (complex tasks). Drafting, analysis, multi-turn synthesis.

10%: Frontier models

GPT-4o, Claude 3.5 Sonnet, Opus. Reserved for complex reasoning, code generation, high-stakes interactions, agentic planning.

This split is arbitrage. You buy intelligence at the exact level the task requires, rather than overpaying for unused reasoning capacity.

How to migrate safely

You can't flip a switch. Migration requires rigor:

1. Log your prompts. Sample 1,000 requests from your most expensive endpoint.
2. Categorize intent. 80% of volume typically comes from 2-3 repetitive task types.
3. Shadow test. Run a cheap model in parallel. Compare outputs programmatically using LLM-as-a-judge or deterministic metrics.
4. Implement routing. Use a lightweight semantic router or rule-based heuristics to direct traffic by task complexity.