Every founder building on top of a language model eventually runs into the same uncomfortable question at the pricing meeting: if the model is the product's engine, who pays for the fuel? Get the answer wrong and you have built a business that loses a little money on every active user, which is a hard thing to grow your way out of.

There are two broad answers. You can resell the model: charge the customer, pay the provider yourself, and try to keep a margin in the middle. Or you can let the customer bring their own key. Bring-your-own-key pricing, BYOK for short, is the quieter of the two options, and for a lot of AI products it is the one that removes the risk you least want to carry.

What bring your own key actually means

In a BYOK model, your product does not hold a model account that it bills against. Instead the customer supplies their own credential, from Anthropic or OpenAI or whoever, and your software calls the provider directly on their behalf. The tokens are billed to the customer by the provider, at the provider's published rates, with no stop in your ledger along the way.

This is not exotic. It is how several of the most widely used AI coding tools already work: the tool is a client, and the intelligence is rented by the user from the model vendor. You are selling the orchestration, the interface, the workflow, and the judgment around the model. You are not selling the model.

The distinction sounds like plumbing. It is actually the whole economic story.

The margin math BYOK avoids

Classic software has beautiful margins because the marginal cost of one more user is close to zero. AI software broke that. When you resell inference, every single call is a real variable cost, so your cost of goods sold climbs back into the picture in a way SaaS founders had almost forgotten about. Public analyses of AI businesses put gross margins in the fifty to sixty percent range, against the eighty to ninety percent that traditional SaaS enjoys. That gap is inference, and it does not shrink as you scale, it grows with usage.

Reselling tokens also hands you a margin you have to defend forever. Model prices move, sometimes down, occasionally up, and your pricing has to track them or your margin quietly erodes. You are now in the business of predicting someone else's price list.

BYOK sidesteps all of it. If the customer pays the provider directly, your variable inference cost is near zero, and you are never exposed to a per-token margin you have to protect. You price the platform layer, the thing you actually build, on its value. The economics of your product stop being a bet on the model vendor's price list. This is the same reason AI product margins look so different from classic software, viewed from the supply side, and it shapes how you price an AI product in the first place.

Why customers often prefer it too

BYOK is not just a hedge for the vendor. It tends to be the honest option for the buyer, and honesty sells.

A customer on their own key pays exactly what they use, at rates they can look up, with no markup baked in that they cannot see. For a light user that is straightforwardly cheaper than a subscription that assumes heavy use. For a heavy user it removes the suspicion, correct surprisingly often, that a reseller is charging a comfortable multiple on top of the real token cost. The bill comes from the model provider, so the customer can reconcile it against a public price list instead of taking your word for it.

It also defuses two objections that kill enterprise deals. There is no lock-in on the inference layer, because the account is theirs. And in regulated settings, keeping the model relationship in the customer's own vendor account is often easier to get through procurement and security review than routing sensitive prompts through a third party's account.

The trade-offs, honestly

BYOK is not free money, and pretending otherwise is how you end up surprised later.

You give up a revenue lever. If inference genuinely carries margin for you, BYOK hands that margin to the provider instead of to you. Some businesses should resell tokens precisely because the markup funds real value they add on top. That is a legitimate model, not a mistake.

You inherit a support surface. Now the customer has to obtain a key, set spending limits, understand rate limits, and occasionally debug a billing problem that is really the provider's. Onboarding gets a step longer, and a confused key setup is a confused first impression.

You lose some control over spend. When the customer owns the account, you cannot hard-cap their bill from inside your product. You can guide, warn, and bound your own dispatch decisions, but the ceiling is on their account, not yours. That has to be communicated plainly, and it pairs naturally with strong cost controls and budgets so a runaway loop is caught before it becomes a bill.

And bundling is harder. "One simple price" is an easier thing to sell than "our fee plus your metered model usage," even when the second is cheaper and more honest. You are asking the buyer to hold two numbers in their head.

The honest limitation

BYOK removes token-margin risk. It does not remove the need to build something worth paying for. If your only value was a thin wrapper over an API, BYOK exposes that, because the customer can now see exactly how little sits between them and the model. That is a feature of the model, not a bug: it forces you to price the part you actually add.

It is also not a universal answer. A product serving non-technical users who will never manage an API key may be right to resell inference and hide the complexity, eating the margin hit as the cost of a simpler experience. The choice is a real trade between honest economics and onboarding friction, and the right call depends on who you sell to.

What BYOK gives you, when it fits, is a business whose margins do not erode every time a model vendor reprices, and a cost story you can tell a customer without flinching.

That is the stance we took with Loopsfinity. You bring your own model account, so inference is billed to you directly with no markup, and we price the platform we build rather than the tokens we did not. The specifics of how we package that are ours, but the principle is not, and it is the one we would want any buyer to hold us to. It is also just one lever in the economics of AI coding agents, which is worth understanding whole before you set a price.