Your pricing is the product

When I started using AI coding tools, I tried Cline and Cursor around the same time. Cline, a VS Code extension, shows a running dollar estimate for the task that updates after every request. I’d watch the number climb and think twice before letting it keep going. Early Cursor was $20 a month for 500 fast requests, then unlimited slower ones, and I never thought about what a request cost. It felt to me like I used Cursor far more because of it, even though the models underneath were often the same.

That’s the thing I keep coming back to. How you charge changes how people use and experience your product, which makes pricing part of the product. Charge per seat and customers get careful about who gets a login, and if you charge per API call they make fewer calls. Charge for storage and people start deleting things, while an unlimited plan lets them stop thinking about it. So decide the behavior you want first, then price for it.

That was always true, and AI makes it hard to ignore. Software spent decades making the next use feel free. Once you paid for the license, the seat or the subscription, using it more cost you nothing, so people used it more. AI brought the cost back, since every request burns real compute, and a lot of products went back to metering with credits, tokens and usage tiers.

Agents make the conflict obvious. The pitch for an agent is to hand it the work and stop worrying about it. A credit meter says to think carefully before you hand it anything, and it makes you feel the cost at the exact moment you’re deciding whether to try something, which is when you want people trying things. When your pricing trains the opposite of what your product needs, that’s a product problem as much as a billing one. Cursor found that out in June 2025 when it switched Pro to $20 of usage at API prices. People ran through it fast, and Cursor ended up apologizing and offering refunds.

Credits exist for good reasons, though. Compute really does cost money, some customers use a hundred times more than others, agents can run for a long time, and unlimited plans attract the heaviest users. I don’t think unlimited is automatically the answer. The better question is where the meter lives. You need one internally. Your customers don’t have to feel it on every request, and your costs can be variable on the inside while the price stays predictable on the outside, by routing to cheaper models, caching and pricing against the average user.

At HubLogix in 2014 we charged by SKUs, vendors and orders. Vendors were the warehouses and suppliers a customer connected, which was the whole point of the product, so every new connection cost them more. In 2015 we moved most of the price onto orders, mostly to move upmarket and make it easier to sell. I think that was the better meter, since it’s simpler and customers paid more when they were selling more. The trade-off was that it taxed their growth and made their bill move every month, when most customers wanted a price they could plan around. We lowered the per-order price as customers grew, and never fully solved predictability, but it worked well enough.

If I were pricing an AI product today, I’d start with the behavior I want, like handing it more work, and design pricing that doesn’t fight it. Then I’d make the economics work inside that design. My guess is the best AI products will end up hiding the meter over time. Compute won’t be free, but managing it will become the company’s job instead of the customer’s.