---
title: "Model Cascade (FrugalGPT) · agent orchestration · Agent Fieldbook"
url: https://agentfieldbook.org/design/orchestration/frugalgpt/
description: "Try the cheapest model first; a confidence/verifier signal decides whether to escalate to a stronger (pricier) model. Queries stop at the first model good enough."
section: "Design · Orchestration"
source: Agent Fieldbook — generated from the published page
---

# Model Cascade (FrugalGPT)

Try the cheapest model first; a confidence/verifier signal decides whether to escalate to a stronger (pricier) model. Queries stop at the first model good enough.

Verified against [arxiv.org](https://arxiv.org/abs/2305.05176)

- [arxiv.org](https://arxiv.org/abs/2305.05176)

## How it works

## At a glance

| Field | Value |
| --- | --- |
| Technique | Cascade |
| Cost lever | Dramatically lower, up to 98% cost reduction reported |
| Latency lever | Lower for easy queries, higher for escalated ones |
| Quality lever | Matches or beats GPT-4 at a fraction of cost on benchmarked tasks |
| Example | FrugalGPT· FORC |

## When to use

Budget-constrained workloads with a long tail of easy queries and a clear signal for when the cheap answer is unreliable.

## Basis

## Verified against

- [https://arxiv.org/abs/2305.05176](https://arxiv.org/abs/2305.05176)

The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.