✓ Last verified: 2026-07-29· FrugalGPT-style cheap→verify→escalate· report an issue →
A model cascade runs a cheap model first, verifies the answer, and only escalates the ones it fails to a premium model — the FrugalGPT pattern. But you pay the cheap model on every request, plus a verifier, plus premium on the escalated share — so it is not free. This prices the real blended monthly cost of a cascade against always-premium and a plain split, and finds the escalation rate where the cascade stops saving.
The verifier runs on every request, not just escalations.
—
cascade / mo
—always-premium / mo
—savings
—break-even escalation
Where the cascade cost comes from
The cheap model and the verifier are paid on the full volume; the premium model is paid only on the escalated share. Always-premium and a plain traffic split are shown for comparison.
Component
Monthly cost
Share
Savings across escalation rates
Escalation rate is the swing factor. This holds everything else fixed and steps it to show where the cascade stops beating always-premium — the break-even row is marked.
Escalation
Cascade / mo
Savings
Verdict
A cheap model first is not a cheap bill by default
The appeal of a cascade is obvious: most requests are easy, a small model gets them right, and paying premium prices for easy work is waste. Run the cheap model first, keep the answers that pass a check, and escalate only the ones that fail — the FrugalGPT result showed this can match a frontier model's quality at a fraction of the cost. The trap is treating the cheap model as the whole cost. You pay it on every request, whether the answer survives or not. You pay a verifier on every request too, because something has to decide what escalates, and that something is usually another model call. And on the share you escalate, you pay the premium model on top of the cheap attempt you already made. Add those up and the picture changes: the cascade saves handsomely when escalation is low, shrinks as escalation climbs, and past a break-even rate it costs more than just calling the premium model on everything — because you have paid the cheap model and verifier for nothing and still paid full premium price. This calculator prices all three legs from your own token counts and rates, compares the cascade to always-premium and to a plain traffic split that never double-pays, and marks the escalation rate where the arithmetic flips. If you route each request to a single model up front instead of escalating, price that with the model router calculator; to stack caching, batching and other levers on top, use the cost optimization calculator; and to compare the two models themselves, see the model comparison.
It starts from the per-request cost of each model: input tokens times input price plus output tokens times output price, divided by a million. The cheap model and the verifier are billed on the full request volume, so their monthly cost is the per-request figure times requests per month. The premium model is billed only on the escalated share, so its monthly cost is the premium per-request cost times requests times the escalation rate. Cascade monthly cost is the sum of all three. Always-premium is simply the premium per-request cost times the full volume; the plain traffic split — shown for comparison — pays the cheap model on the non-escalated share and the premium model on the escalated share, never both. Savings is one minus cascade over always-premium. The break-even escalation rate is one minus the ratio of the cheap-plus-verifier per-request cost to the premium per-request cost: below it the cascade beats always-premium, above it the cheap-first attempt plus verifier is dead weight and going straight to premium is cheaper. The model assumes escalated requests use the same token counts on both models; if your premium calls are longer, raise the premium token counts to reflect it.
Frequently asked questions
What is an LLM cascade and how does it save money?
It answers each request with the cheapest model that can handle it: a cheap model runs first, a verifier decides if the answer is good enough, and only the failures escalate to a premium model. Most requests are easy, so the cheap model clears them at a fraction of the cost. The saving is real, but you still pay the cheap model and verifier on every request plus premium on the escalated share.
How is a cascade different from routing between two models?
A router picks one model up front, so each request pays for one model and never double-pays — but a misroute ships a wrong answer. A cascade runs the cheap model on everything and escalates failures, so escalated requests pay cheap plus verifier plus premium, but every hard request gets a proper second attempt. This tool shows both costs at the same escalation rate.
At what escalation rate does a cascade stop saving?
At the break-even escalation rate — one minus the ratio of cheap-plus-verifier per-request cost to premium per-request cost. Below it the cascade beats always-premium; above it you have paid the cheap model and verifier for little benefit and still paid full premium, so calling premium directly is cheaper. The cheaper your small model and verifier, the higher escalation can climb before the cascade turns into a loss.
What counts as the verifier cost and why does it matter?
The verifier is whatever decides to ship or escalate — a small judge model, a self-consistency pass, or an embedding threshold — and it runs on every request, so a per-call cost that looks trivial adds up at volume. It also sets the floor: even with zero escalation you pay the cheap model plus the verifier on the whole volume. This calculator treats it as a first-class input rather than rounding it to zero.