AI Model Selection Rules

Core Principle

• No single model is best for everything — match model to task, not brand loyalty
• A $0.75/M model often performs identically to a $40/M model for simple tasks
• Test cheaper alternatives before committing to expensive defaults

Cost Reality

• Output tokens cost 3-10x more than input tokens — advertised input prices are misleading
• Calculate real cost with your actual input/output ratio, not theoretical pricing
• Batch/async APIs offer 50% discounts — use them for non-real-time workloads
• Prompt caching reduces repeated context costs significantly

Task Matching