We Tested Stripe, Coinbase, and Circle on AI Agent Payments. Here Is the Math

Claim 1 · empirical
Under a standard per-transaction execution model, Coinbase and Base still incur individual micro-fees on every hop; across 1,000 continuous requests those Layer-2 gas fees accumulate between $5.00 and over $40.00 in extra overhead, an extra cost burden of 12.5% to over 100% on a $40.00 workload.

What would break it: Measured Base Layer-2 gas for 1,000 continuous micro-transactions totals less than $5.00, putting overhead below the stated 12.5% floor.

Claim 2 · empirical
At the 100-request scale ($4.00 baseline data cost), Stripe fixed fees add $30.00 of overhead, totaling $34.00.

What would break it: Measured Stripe overhead on 100 individual micro-requests at the same baseline comes in materially below $30.00.

Claim 3 · empirical
At low volume, Circle shows a higher percentage overhead initially (roughly 12.5% to 30% on a 100-request test) because the initial deposit cost is spread across fewer requests.

What would break it: Measured Circle Nanopayments overhead on a first 100-request run from a freshly funded balance falls below 12.5%.

Claim 4 · mechanism
The funded Circle balance supports future calls over time without resetting the entry fee, so as the user continues to utilize the system the effective overhead drops toward the 1% floor, making the architecture efficient for long-cycle automated operations.

What would break it: Continued usage shows the entry or settlement fee recurring so that long-run effective overhead stays materially above the 1% floor despite balance reuse.

Claim 5 · thesis
If every micro-query an AI agent makes triggers an isolated billing event, the payment rail costs more than the data itself; legacy payment rails collapse when an agent operates at high request velocity.

What would break it: Production agent workloads execute thousands of continuous sub-dollar requests on standard per-transaction payment rails with total payment overhead at or below the underlying data cost.

Claim 6 · mechanism
Circle Nanopayments utilizes a deposit once, authorize many, settle in batches model: the user funds an operating balance with a single initial transaction, paying a one-time network settlement fee of roughly $0.50 to $2.00, and from that point the agent signs off-chain cryptographic authorizations for each data request, so the requests execute without individual network gas fees.

What would break it: Circle Nanopayments is shown to charge a per-request network or gas fee, or to require a new on-chain settlement for each individual request.

Claim 7 · empirical
Stripe x402 / Machine Payments Protocol (MPP) relies on a fixed-fee floor of roughly $0.30 per transaction; for 1,000 individual micro-requests at a $0.04 per-request baseline ($40.00 of data), that adds $300.00 in transaction fees, turning the run into $340.00 and making the payment processing layer seven times more expensive than the intelligence being purchased.

What would break it: Stripe x402/MPP per-transaction fees on micro-requests measure materially below the roughly $0.30 floor, or a 1,000-request run at the same baseline totals materially less than $340.00.

Claim 8 · empirical
On the 1,000-request test, Circle Nanopayments keeps total payment overhead between 1% and 5%, so the final workload cost lands at $40.50 to $42.00 on the $40.00 data baseline and the payment rail never eats the product margin.

What would break it: A 1,000-request run through Circle Nanopayments at the same baseline totals more than $42.00 in combined data and payment cost, putting overhead above 5%.