The more Duplex learns, the less each decision costs.

A model has no way to get cheaper at your work, because your work doesn’t change it. It learned the world before it was released, and every call starts from there, plus whatever the prompt brings along. A hard case nobody has seen before gets its full reasoning, which is what you want. So does a question your team answered months ago, at the same price, and much of what reaches your AI is that second kind.

Teams fight the bill in sensible ways. A shorter prompt trims each call. A smaller model lowers its price, and for easy work that is a fair trade. A cache answers a repeat for almost nothing, and when the repeat is exact it is hard to beat. All of them help, and none of them changes the deal: every call the model makes is still paid for, and a cache only knows what looks the same, so it gives a look-alike the old answer and goes on giving it after your rules change.

Duplex learns instead. When your team confirms a result, Duplex learns the decision behind it, and the next request like it is answered by Duplex itself, with no model call. On a stream of 3,000 decisions, a learned decision came back in About 97 ms. A model call took About 2,231 ms, and you paid for each one. For a customer waiting in a chat, that is the difference between an instant reply and a pause they notice.

So each decision your team confirms comes off the bill from then on. The policy questions, routing choices and familiar exceptions that make up so much of the traffic move to Duplex one by one, and your AI spends its effort where its price is worth paying: on what is new, briefed with what Duplex knows about the case. Once the answer to that new case is approved, Duplex takes it on as well.

Cost and accuracy usually pull against each other. To get more answers right, you pay for a bigger model, a longer prompt or a second check, and to spend less you accept more mistakes. Anyone who has run AI in production has turned that dial both ways, looking for a setting where both get better.

With Duplex the two move together, in the direction you want. Over those 3,000 decisions, with five rule changes along the way, its accuracy rose from 85.0% to 94.6% while the cost of each decision fell by about two thirds. Each decision it learns is answered the way your team decided it, with no call to pay for, and a correction made once applies every time after. So every result it learns improves both.

Set against the same model working alone on the same decisions, Duplex made 90.3% fewer model calls at 84.8% lower cost, and it was right far more often: 94.6% against 59.8%. The gap was widest after the rules changed. On the decisions those changes affected, Duplex was right 92.5% of the time and the model alone 2.3%. The model with a cache managed 42.6%, and a cheap answer is no saving when it follows the old rule.

The result held at every setting tested. Across eight of them, Duplex stayed between 94.1% and 94.8% accurate, at about a seventh of what the model alone costs, while the best a cache reached was 77.3%. And the curve doesn’t have to start at the top: pretrain Duplex on what your team already knows, and the savings begin on the first day.

With a rented model, use is what you pay for, and every success makes the bill bigger. With Duplex beside it, use is what makes the work cheaper: every result your team confirms makes the next decision like it faster, cheaper and more likely to be right. Keep paying your AI for what is new and hard, the work it does best, and let everything your team has already settled cost less every month.