Cheap LLM Routing: Flash vs Frontier Models for Everyday Tasks

Cheap LLM routing is the practice of answering the easy 90% of requests with fast, flash-class models and reserving frontier models for the hard 10%. The split alone typically halves your bill, and you can tune it with a quality threshold.

This guide is the practical how-to: how to define easy vs hard, where to set the threshold, and how to measure before you trust it.

Define 'easy' for your workload

• Classification and extraction are usually easy: structured, short, low risk.
• Summarization is easy when the source is short and the format is fixed.
• Creative, multi-step, or code-heavy tasks are usually hard.
• Reputation defines difficulty better than length does.

The split and the threshold

1. Route clearly easy traffic to a flash model by default.
2. Escalate when a cheap model signals low confidence (score thresholds, schema failures).
3. Add a fallback: retry failures automatically on the frontier model.
4. Monitor quality complaints vs cost saved each week.
5. Adjust the threshold until the quality-cost balance fits.

Expected savings

On mixed workloads, 40-60% savings is a safe estimate when 80-90% of traffic is easy. If your traffic is mostly hard, routing saves little — which is useful information by itself.

FAQ