Model Cascade
A cost-optimization pattern where a cheaper, faster model handles the first attempt at a task, and a more capable (but slower and more expensive) model is invoked only when the first model fails a confidence threshold or quality check. Cascades typically cut total cost 60-90% on workloads dominated by easy cases (a large fraction of support tickets, document classification, simple coding tasks), because most queries never reach the expensive model.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A customer-service AI agent fronts every query with a small model (Haiku, GPT-5-mini) that handles 70% of requests directly. For the 30% where the small model is uncertain—measured by output confidence, retrieval scores, or a self-check—the request escalates to a frontier model (Claude Opus 4, GPT-5). Average per-ticket cost drops from $0.18 to $0.06 with no measurable quality loss on the cascade-handled cases.
Frequently asked questions
- When does a model cascade NOT pay off?
- When the task distribution is mostly hard cases (the cheap model fails on 60%+ of inputs), when latency budget is tight (cascading adds a second model call when escalating), or when the cheap model produces *plausibly wrong* answers the system can't easily detect as wrong. The cascade only wins when easy cases dominate AND the cheap model's failures are reliably caught.
- How do you decide when to escalate?
- Common signals: low retrieval scores, the cheap model's own self-rated confidence, mismatched outputs across two passes of the cheap model, or a fast LLM-as-judge check on the cheap model's output. Threshold tuning is empirical—start with a conservative threshold (escalate often), measure quality and cost, then tighten until you hit the cost-quality target.