What a subscription is actually doing#
It is a risk trade dressed as a price.
The seller takes usage risk: a subscriber who calls ten times and one who calls ten thousand times pay the same, so the seller carries the variance. In exchange the seller gets predictable revenue and a relationship it can forecast against.
The buyer takes commitment risk: it pays for the month whether or not it uses the month. In exchange it gets a predictable bill and no per-call anxiety.
That trade works when both parties expect to be around. It stops working on both sides when the buyer is software that found you a moment ago and wants one call.
What breaks when the buyer is an agent#
| Subscription assumption | Still true for an agent buyer? |
|---|---|
| The buyer can commit to a month | No. It may exist for ten minutes |
| The buyer will be back | Unknown, and usually not |
| A human evaluates the price | No. A program compares numbers |
| Sign-up is a small one-time cost | No. It is the thing that stops the purchase |
| Usage is roughly predictable | No. It is bursty by construction |
The third row is the one sellers underestimate. A human weighing a subscription considers switching cost, familiarity and the hassle of changing. A program comparing two endpoints that return the same data considers price and latency. It has no loyalty to spend and no inertia to overcome, which means a bundle whose value was partly the buyer's reluctance to re-evaluate is worth less than it was.
Does per-call pricing just lose you money?#
Not necessarily, and the comparison people make is usually wrong.
The wrong comparison is per-call revenue against subscription revenue from the same buyer. By that measure the subscription usually wins, because it was priced to.
The right comparison includes the buyers you do not currently have. An agent that abandons at the sign-up form generates zero, not a subscription. For a product whose addressable demand is increasingly software, the question is not whether per-call is cheaper per buyer but how much of the market the sign-up is excluding.
Note
The honest version of this: per-call pricing is usually worse revenue per existing customer and better revenue per unit of demand you were not reaching. Whether that trade is positive depends on how much agent traffic your category actually has, which you can measure before deciding.
What does the hybrid look like?#
Most AI products should end up with three tiers rather than two.
- Free, limited. For evaluation, by humans and by crawlers. Rate limited, no payment path. Its job is to let a buyer see your interface before deciding.
- Per-call, anonymous. x402, no sign-up, priced per request. Its job is to convert a buyer who arrived thirty seconds ago and would otherwise leave.
- Committed, identified. Keys, invoices, support, volume pricing. Its job is the relationship with buyers who will integrate and stay.
The connection between 2 and 3 is where the value is. Let a payment mint a key: the buyer pays anonymously once, gets a credential, and crosses from stranger to customer with no form and no human on either side. Keys against x402 covers that pattern in detail.
How should the per-call price relate to the subscription?#
Above the subscription's effective per-call rate, always.
If your plan is a hundred dollars for a hundred thousand calls, your effective rate is a tenth of a cent. Pricing per-call at a tenth of a cent gives committed buyers a reason to leave the plan, and you have cannibalised predictable revenue to win none.
Per-call should carry a premium, because it is a premium product: no commitment, no minimum, instant access, no relationship. Buyers who want that convenience can pay for it. Buyers who want the lowest unit price can commit, which is what the plan is for.
What about metered billing without a subscription?#
Usage-based billing with an account is a middle ground and a genuinely good model for identified buyers. It keeps the relationship and removes the commitment.
What it does not remove is the account, which is the thing the agent cannot complete. Metered billing is a better subscription, not a substitute for anonymous payment. If the problem you are solving is "my buyer cannot sign up", metered billing does not solve it.
Frequently asked questions#
Will per-call pricing make my revenue unpredictable?#
Yes, that part is real. It is the cost of the trade. Most sellers run per-call alongside a subscription book rather than instead of one, so the predictable base remains.
How do I price a call I have never sold individually?#
Start from your effective per-call rate under your existing plans, add a premium for the convenience, and compare against what comparable endpoints charge. /services shows the listed price distribution across indexed endpoints, which is a better anchor than guessing.
What if my value is in the corpus, not the call?#
Then per-call pricing fits badly and you should say so rather than forcing it. Products whose value is a session, a relationship or an accumulated state are subscription-shaped, and no payment protocol changes that.
Does per-call pricing work for inference?#
Yes, with a scheme that handles variable cost. Exact against upto covers that specifically, since token-metered output is the case fixed pricing handles worst.
Can I test demand before committing?#
Yes, cheaply. Add a paid path to one endpoint, publish it in a catalogue, and see whether anything pays. That experiment costs a day and answers the question better than reasoning about it.
Where to check this yourself#
Before pricing anything, look at what the market is actually paying. /services lists indexed endpoints with their listed prices and observed activity, and /transactions shows settled amounts rather than advertised ones. The gap between the two is the most useful number in the exercise.