4.1.2 · Lesson 2 of 16
Cost-Effective Model Selection & Model Cascading
Match model capacity to query complexity instead of routing all traffic to top-tier reasoning models.
WHAT
Match model capacity to query complexity instead of routing all traffic to top-tier reasoning models.
WHY
Top-tier models can cost 10–50x more per token than lightweight models (e.g., Claude 3 Haiku / Titan Text Express).
HOW
Model Cascading (Tiered Intent Routing): classify with a fast low-cost model first; escalate when confidence is low or reasoning is required.
User Request
↓
Lightweight Intent Classifier / Fast FM
├── Simple Query → Return immediately
└── Complex Query → Cascade to High-Reasoning FMKey points
- Model cascading / tiered routing
- Lightweight classifier first
- Escalate complex queries
- Haiku vs Sonnet cost gap
- Confidence threshold routing
Exam focus
Route simple queries to cheap/fast models and cascade only complex work to high-reasoning FMs — do not send everything to the largest model.