Not Every Task Needs the Best Model

What this chapter covers

Frontier models are expensive because they are general. Most enterprise tasks are narrow. This chapter is about routing — using cheaper, faster, or specialised models for the majority of calls and reserving the best model for the moments that need it.

Key takeaways

• Model routing patterns: cascade, ensemble, and tiered fallback
• How to classify a task by difficulty before you spend a token on it
• When a small fine-tuned model beats a large general one
• How routing changes latency, accuracy, and cost together
• A worked example: cutting cost 70 percent on a support workflow