The $0.03 inbox vs the $0.17 inbox —
and not a single ticket apart.
Four AIs worked the same 25-ticket support queue for CloudCanvas — refunds, a legal threat, a GDPR deletion, replies owed in Spanish, German and Japanese. Every one cleared all 25. We stripped the labels. Below is the hardest ticket in the stack: a legal threat that must be escalated, never argued. Read the four replies, guess which run cost 5.7× the most, then reveal the receipts.
A four-way tie on quality. A 5.7× spread on the bill.
Every model cleared all 25 tickets — the refund denials, the legal escalation, the GDPR deletion, the German double-charge, replies owed in three languages. Nobody fabricated an invoice. Nobody caved on policy.
So the answer isn't in the replies — you can't tell them apart. It's on the invoice. One inbox cost six cents on the cheapest route and seventeen on the priciest, for work a grader scored identically. Most teams are hard-coded to the expensive end and never notice.
Tickets cleared, out of 25
A ticket clears only if every criterion is met — right policy call, no fabricated facts, correct language, no forbidden exclamation marks. Four bars, all the way to the wall.
Same traps. Same blind labels. Read them side by side.
Ticket #2 is a refund demand that policy says to deny — the tempting wrong answer is to invent an exception. Ticket #6 must be answered in Spanish with the exact export path. Switch tabs; the labels stay withheld.
The verdict
All four cleared the inbox. Every refund call correct, every policy trap held, three languages handled, no fabricated invoices, no forbidden exclamation marks. On the grader it's a 25/25 four-way tie — which is exactly why we hid the labels.
So the only thing that ranks them is the bill, and the bill ranges 5.7×. Four inboxes you genuinely can't rank by quality, priced from three cents to seventeen.
Pick by brand and you're gambling with your budget. Pick the cheapest that still clears the bar — which is all routing is — and you keep the result while cutting the bill to a fraction. "We'll send you to whatever's cheapest that works" is a sentence no single lab can say. We can.
All four cleared 25/25 — so this is a pure price race. One full 25-ticket run, one shot each.
| Reply | Cost / inbox | vs cheapest |
|---|---|---|
| Reply CCHEAPEST | $0.0303 | 1× |
| Reply A | $0.0820 | 2.7× |
| Reply D | $0.0939 | 3.1× |
| Reply BPRICIEST | $0.1719 | 5.7× |
Four leading models ran this inbox — Claude Fable 5, Claude Opus 4.8, GPT-5.5, and one more. We're not telling you which is which, because you couldn't tell either. That's the point.
The answer isn't picking a model. It's routing to the cheapest one that clears the bar.
Every model in this contest cleared all 25 tickets — including the one that did it at 1/5.7 the cost of the priciest. Lock yourself to a single frontier vendor and you pay the top of that range on every ticket, forever. Pareto sends each request to the cheapest model that still passes your quality bar, proves it against the grader, and gives you one endpoint. You stop choosing. You just stop overpaying.
One shot each
Identical 25-ticket inbox, same system prompt, no cherry-picking. Each reply is exactly what the model returned.
Graded on the rules
A ticket clears only if every criterion is met — right policy call, correct language, no fabricated facts, no forbidden exclamation marks. All four cleared 25/25.
Real bills
Fable/Opus: delivered output chars ÷ 4 × published rate ($50/M out Fable, $25/M Opus). GPT-5.5 and the GLM route: real usage tokens at their at-cost gateway rates.
Labels withheld on purpose
We don't map model to letter. If quality were the differentiator you'd spot the expensive one. You can't — so it isn't.
Ran 2026-07-04 · Two models ran as Claude agent runs via the Anthropic API; GPT-5.5 and one more ran through our at-cost gateway. Same prompt, one shot each.
Send us your AI bill.
We route every call to the cheapest model that clears your quality bar — and prove it nightly. What you spent, what you'd have spent, receipts attached.