The Model Selection Problem

Most people use one model for everything. This is a mistake.

The Problem

Using Claude Opus for everything:

• Expensive ($15/M input, $75/M output)
• Slow (more reasoning = more latency)
• Overkill for simple tasks

Using only cheap models:

• Poor quality for complex reasoning
• More errors and hallucinations
• Misses nuance

The Solution: Model Hierarchy

Different tasks need different models:

Real Example

My setup:

• Main agent (coordinator): Opus — needs full reasoning
• Background tasks: Sonnet — good enough, cheaper
• Quick lookups: Haiku — fast, very cheap
• Cron jobs: Sonnet — isolated, no need for Opus

This cuts costs by ~60% without quality loss.

The Key Insight

Match the model to the task, not the other way around.

Complex task → Smart model Simple task → Fast model Background task → Cheap model

The goal is optimal cost-quality tradeoff, not maximum intelligence.