Most teams pick a prompt style by feel. The answers look shaky, so someone adds a few examples or writes "think step by step", and the answers get better. What nobody checks is the bill. A bigger prompt gets paid for on every request, every day.
So I tested it. Same model, same questions, three prompt styles, and I counted every cent.
The short version is that the most expensive prompt won one job and came last in the other. Which one you should pay for depends on the job, and on what a wrong answer costs your business.
The three prompt styles
Zero-shot just describes the task and lists the answers the model can pick from. Few-shot adds a handful of solved examples for the model to copy. Chain-of-Thought asks the model to work through the problem step by step before it answers.
Chain-of-Thought costs more for a simple reason. The model writes out its thinking, and on this model, writing costs about 8x more than reading.
Here's what each one looked like for the support ticket job.
The two jobs
The first job is routing support tickets. Every customer message goes to one team: Billing, Technical, Refund, Account or Urgent.
The second is tagging products for an online store, as Electronics, Accessories, Home, Media or Outdoor. This store has its own habits. A Kindle goes under Media. A charging cable goes under Accessories, even though it's electronic. Those habits aren't written down anywhere, they're just how past products were tagged. A lot of real stores work like that.
Each job got 40 test items, and half of them are hard on purpose. They try to trick or pressure the model.
I wrote all the test items and the right answers myself. Everything ran on Gemini 3.5 Flash Lite through OpenRouter, at $0.30 per million input tokens and $2.50 per million output tokens. I turned the model's own built-in reasoning down to minimal, so any extra thinking had to come from the prompt.
Support tickets, where thinking paid off
Chain-of-Thought routed all 40 tickets correctly. Few-shot got 38, and Zero-shot got 31.
Zero-shot mostly went by the tone of the message. A calm "no hurry, a former employee still has admin access" went to Account, when it's a security problem. A dramatic "catastrophic emergency" about a slow app went straight to Urgent.
Going from Few-shot to Chain-of-Thought got 2 more tickets right out of 40, and cost 4x as much.
Product tags, where thinking lost
Few-shot got 38 of 40 right. Chain-of-Thought got 27, barely better than Zero-shot's 26, and it cost 6x more than Few-shot.
It thought about the Kindle the way most people would. It's a battery-powered device with a screen, so it must be Electronics. That's a sensible answer, and it's wrong for this store. The rule that e-readers go under Media only lives in past tags, and thinking harder can't find a rule you've never seen. Few-shot had one example, a Kobo e-reader tagged as Media, and that was enough.
All 13 of Chain-of-Thought's mistakes were the same kind. It filed something under Electronics that this store keeps somewhere else.
When the right answer depends on your company's own rules, examples teach those rules. Reasoning doesn't.
What a wrong answer costs you
Looking only at the charts, it seems simple: Chain-of-Thought for tickets, Few-shot for products. That skips the question that actually decides it, which is what a mistake costs you.
In ticket routing, switching from Few-shot to Chain-of-Thought costs an extra $273 per million tickets and fixes about 50,000 more of them. That's roughly half a cent per fixed ticket.
If a misrouted ticket just waits a few minutes until an agent moves it, half a cent per fix may not be worth paying. If it's a data leak sitting in the wrong queue overnight, half a cent is nothing. Try it with your own numbers.
When a mistake costs nothing, the cheapest prompt always comes out ahead. For tickets, Chain-of-Thought takes over once a mistake costs more than about half a cent. For products it never does, because there it costs more and gets fewer right.
Where this test is thin
It's 40 items per job, so a small gap like 95% vs 100% is only two tickets. It's one cheap model, and a bigger one might pick up the store's habits without examples. I wrote the test data myself. Run something like this on your own data before you bet a budget on my numbers.
The business call
If this landed on my desk as a budget decision, I wouldn't ask which prompt is best. I'd ask what one wrong answer costs us. Put a number on it, even a rough one. A misrouted ticket costs agent time and a slower reply. A product in the wrong category costs sales, because shoppers don't find it.
Next comes volume. A gap of a few hundred dollars per million requests doesn't matter at 10,000 requests a month. At 50 million, it's a line item someone in finance will ask about.
Then it's the same math as any other purchase. Work out what each extra right answer costs. Here it was half a cent for tickets, and for products the expensive option bought fewer right answers. If the extra accuracy costs less than the mistakes it prevents, pay for it. If it doesn't, keep the money.
And check again when anything moves. A price change, a new model or a jump in volume can flip the answer, so treat the prompt style like any other cost you review each quarter.