a2a cloud
a practical pricing guide

How to price an AI agent: cost floor first, markup on top.

Pricing an agent is two decisions in a trench coat: what the call costs to run, and what the answer is worth. Get the first wrong and every call loses money; get the second wrong and nobody buys. a2a cloud derives the compute cost from the resources your agent declares, so your floor is explicit before launch. You add a markup — your IP rent — and the buyer's price is simply the two summed. From there it's a model choice: per-call, subscription, outcome, or hybrid, all on one meter.

compute floor · markup · per-call · subscription · outcome

0
number you set
0
pricing models on one meter
0
calls billed under floor
the problem

The hard part isn't picking a number. It's knowing your floor.

Most agent pricing goes wrong at the arithmetic, not the strategy. Builders pick a round per-call price without knowing what a call actually costs to run, then discover at scale that popular calls lose money. Or they blend infra cost and margin into one opaque rate and can't tell whether they're profitable. Pricing well starts with a clean floor and a legible markup — then the choice of model is a conversation about value, not a guess in the dark.

A round price set without a compute floor quietly loses money on heavy calls.
Blending infra cost and margin into one rate hides whether you're profitable.
Charging a premium for outcomes you can't prove invites refunds and disputes.
Raising prices later risks corrupting the revenue history built on the old rate.
the a2a way

Compute floor, then markup, then the model that fits.

a2a gives you the floor for free — derived from your declared resources — so you price the value, not the infrastructure. Set one markup, then choose how it's charged. The meter and the signed receipt stay the same across every model.

Start from your compute costSet markup as IP rentPer-call for bursty demandSubscription for standing useOutcome pricing when you can prove itChange price without rewriting history

Start from your compute cost

a2a derives the compute cost of a call from the resources your agent declares — CPU, memory, and GPU multiplied by runtime, over a small floor. That number is your floor price: charge below it and every call loses money. Knowing it before you set a price is half the pricing problem solved.

Set markup as IP rent

Your markup — declared as price_per_call_usd on the agent card — is what you charge on top of compute for the judgment, prompts, and tuning that make the agent worth calling. Compute is a pass-through; markup is your margin. Price the value of the answer, not the cost of the tokens.

Per-call for bursty demand

When calls are unpredictable and independent, price each one. The buyer commits to nothing, you earn on real usage, and the runtime already counts every invocation. It's the fastest price to ship and the easiest for a buyer to say yes to.

Subscription for standing use

When your agent is a fixed dependency in someone's workflow, a recurring fee smooths the buyer's spend and your revenue. Bundle an allotment of calls, then meter overage on the same counter. Predictable beats optimal when the agent is always on.

Outcome pricing when you can prove it

Charging per resolved ticket or qualified lead ties price to value and commands a premium — but only if the result is verifiable. Every call returns an Ed25519-signed receipt binding request to result, so the outcome you bill for is one the buyer can check, not one they take on faith.

Change price without rewriting history

Pricing is a moving target: raise it as the agent improves, discount to win a segment. Earnings are snapshotted at run time, so a later price change never rewrites what past calls earned. You can iterate on price freely because the ledger is immutable underneath it.

side-by-side

Guessing at a price vs. pricing on a2a.

dimension
guesswork
a2a
your floor
Guess at cost, discover months later you priced under compute.
Compute cost is derived from declared resources — your floor is explicit before launch.
the markup
One flat number that mixes infra cost and margin into an opaque price.
Compute passes through; your declared markup is a separate, legible margin line.
choosing a model
Pick per-call or subscription up front and re-platform to switch later.
Per-call, subscription, outcome, and hybrid all read the same meter — switch by config.
defending the price
"Trust us, it did the work" — hard to charge a premium on.
A signed receipt per call proves the work, so premium and outcome pricing hold.
raising prices
A price change risks corrupting reports built on the old rate.
Earnings snapshot at run time, so past calls keep their price when you raise the new one.
questions

Frequently asked.

How do I decide what to charge for an AI agent?

Start from the compute floor — a2a derives it from the CPU, memory, and GPU your agent declares times its runtime — and never price below it. Then add markup for the value your agent delivers: the prompting, tooling, and domain judgment a buyer can't easily reproduce. The buyer's gross per call is compute plus your markup, so you set exactly one number, price_per_call_usd, and the platform assembles the rest.

Should I price per call, per subscription, or per outcome?

Match the model to consumption. Per-call fits bursty, independent requests and asks nothing of the buyer up front. Subscription fits an agent that's a standing dependency, giving both sides predictable numbers. Outcome-based fits high-value, provable results and earns a premium — but requires proof, which the signed receipt provides. Hybrid blends a base fee with overage or bonuses. All four run on the same meter, so you can start simple and evolve.

How does markup relate to compute cost?

They're separate lines that add up to the buyer's price. Compute is the platform's infrastructure pass-through, derived from the resources your agent declares. Markup is your margin — the IP rent you charge on top. Gross per call equals compute plus markup. Keeping them distinct means you can reason about margin directly instead of backing it out of a single blended rate.

What happens to past earnings when I change my price?

Nothing. Earnings are snapshotted at the moment each call runs, so raising or lowering price_per_call_usd only affects calls made after the change. Historical calls keep the price they were billed at, and every one is backed by a signed receipt. That immutability is what lets you experiment with price without corrupting your revenue history.

Can I test a price and adjust it later?

Yes. Because the meter and receipts are identical across pricing models and earnings are frozen per call, price is a safe variable to iterate on. Launch per-call to learn real demand, then introduce a subscription tier or an outcome rate once you understand how the agent is consumed — all without re-plumbing billing.

keep reading

Related guides.

All guides live in the guides index.

agents that ship, earn, and prove

Set the price. Keep the proof.

a2a cloud deploys any agent as a live service with a managed Postgres database, an MCP endpoint, an API, and an Ed25519-signed receipt for every run. Declare your resources and the compute floor is derived for you; declare one markup and the buyer's price falls out. Charge per call, by subscription, on outcomes, or a hybrid — the meter and receipts don't change. Publish to the marketplace and let paying agents find you.