AI Agents Are Paying in Crypto: 138 Million Micropayments Through x402 on Base

October 25, 2026 MM Crypto Team

AI Agents Are Now Paying for Things in Crypto

Something strange and new is happening on the internet: software is buying things. Not people using software to buy things, but AI agents themselves — spending cryptocurrency to purchase the digital services they need. A Chainalysis study counted more than 138 million transactions through x402 on Base, Coinbase’s layer-2 network, over the 16 months ending in September — a market for automated payments that barely existed two years ago.

The logic is mundane. AI agents need computing power, live web data, and coding tools to complete their tasks. Developers give them budgets, and the agents spend them like a corporate card — except the purchases are so small and so frequent that card processing becomes uneconomic. A payment of a fraction of a cent cannot carry a 30-cent card fee, so the agents pay in crypto instead.

That shift gives cryptocurrency something it has long promised and rarely delivered: a genuine, everyday use case that does not depend on speculation. If machines become economic actors, the payment rails they use will matter enormously.

What Is x402 and How Does It Work

x402 is an open payment protocol built around a piece of internet history: the HTTP 402 status code. When the web was designed, 402 was reserved for “Payment Required” — a slot left empty for a future in which machines could pay for content. For decades it sat unused: the web had no fast, low-cost, programmable money layer. Stablecoins and low-fee layer-2 networks such as Base finally made it practical, and Coinbase introduced x402 to fill the gap.

The flow is deliberately simple. An agent requests a resource; the server replies with a 402 response stating the price and where to pay. The agent signs a payment authorization from its wallet, retries with proof of payment, and the server verifies it onchain and returns the resource. There are no accounts, no API keys, and no subscriptions. Payments settle in stablecoins, with USDC on Base as the default, inside a single request cycle.

That is what makes x402 different from traditional rails. Cards need a merchant account, a processor, and fees built for human-scale purchases. x402 needs only a wallet and a price, which is why it works for transactions worth a hundredth of a cent — amounts no card network would ever touch.

138 Million Transactions on Base

The headline number comes from a Chainalysis study: more than 138 million transactions moved through x402 on Base over the 16 months ending in September — an average of roughly 8.6 million a month through a single payment protocol on a single layer-2 network. For a standard that only began gaining traction in the past two years, the pace is remarkable.

The timeframe matters. Sixteen months ending in September captures the period in which AI agents moved from laboratory demos to production tools — coding assistants, research agents, and automated workflows that businesses actually pay for. As more agents came online and more services priced themselves per request, the payment counts compounded.

It is also worth noting what the number does not say. A transaction count is not a dollar volume, and 138 million payments worth fractions of a cent add up to far less than the same count of card transactions. But the count is the right metric here: it measures how often machines choose to pay — and how often the infrastructure works well enough for them to do it again.

The Micropayment Economics: Cents and Fractions of Cents

The most revealing figure in the Chainalysis study is the size distribution: more than 87% of payments were between a hundredth of a cent and 10 cents. A conventional card transaction carries fixed fees — often around 30 cents plus a percentage — so a one-cent purchase would cost thirty times its own value in fees alone.

A Coinbase Institute paper published on October 7, 2026, made the arithmetic explicit: a $0.30 fixed fee represents a 30,000% overhead on an API query priced at $0.001. In a test case on Base, a transaction of 0.01 USDC settled in two seconds with a network cost below $0.001 through x402. Crypto fills the gap not by being slightly cheaper than cards, but by operating where cards cannot function at all.

This is why micropayments have been a thirty-year unfulfilled promise of the internet. Publishers, data providers, and API operators have always wanted to charge per article, per query, or per computation. Stablecoins on low-fee layer-2 networks finally supply the infrastructure, and AI agents — which consume services in tiny, frequent increments — are the first customers who genuinely need it.

What AI Agents Actually Buy

So what are the agents spending their budgets on? Developers give agents funds to acquire computing power, live web data, and coding tools automatically. An agent working on a task buys the inputs it needs, one at a time, as the task unfolds. It might purchase fresh market data, obtain a model response, and buy a code review — paying separately for each input rather than subscribing to a bundle it may not use.

An AI agent does not know in advance exactly which data it will need; it discovers the need as it reasons through a problem. A subscription assumes predictable usage, but agent usage is inherently unpredictable — a research task might need three data lookups or three hundred. Paying per request matches cost to actual consumption, which is why developers prefer giving agents a budget over buying subscriptions.

Developer tools that charge per request — web scraping, browser automation, image generation — dominate early x402 volume, because they are the tools agents reach for most. The pattern suggests a new kind of software business is forming: services priced per unit of work, sold to customers who are programs rather than people.

A Machine-to-Machine Economy Takes Shape

Publishers are already moving in this direction. Coinbase and Amazon Web Services now let publishers using CloudFront and the Web Application Firewall charge AI agents for content access through x402, with payment handled inside a single request cycle — infrastructure AWS says powers roughly a quarter of the internet. When a quarter of the web can charge agents per request, the addressable market for machine payments stops being theoretical.

The implications ripple outward. If agents can pay for data, data owners can monetize directly instead of relying on advertising or subscriptions. If agents can pay for compute, spare capacity can be sold by the second. The internet’s business model has always been built around human attention; a machine economy is built around machine consumption, and x402 is the cash register.

The Caveat: How Autonomous Is It Really

Before declaring a robot economy, the reporting adds a necessary caveat: the scale of genuinely autonomous commerce remains harder to establish than headline payment counts suggest. A transaction through x402 is a machine payment, but not necessarily an autonomous one. Many are initiated by developers testing systems or running supervised workflows with human-set budgets. The payment is automated; the decision behind it often is not.

Independent analysis sharpens the point. TRM Labs found about $52.7 million across roughly 198.9 million x402 settlements since May 2025, but only 0.6% to 7.5% of that activity looked genuinely agentic after filtering. A separate examination by a16z crypto found a widely cited $24 million figure for agent payments over 30 days fell to about $3 million on inspection, and roughly $1.6 million after removing wash trades. The rails are real and growing, but demand is still catching up.

That gap is normal for infrastructure. Payment networks are built before the commerce that uses them, and the honest reading of the 138 million transactions is that developers are building and testing at scale — not that robots have taken over the economy. The question is whether genuine autonomous spending grows into the rails, and the next year of data will tell.

What This Means for Crypto Markets

For crypto markets, the x402 story lands differently from most AI-crypto narratives. It is not about a token, a fund, or a price prediction. It is about usage — real transactions, on a real network, paying for real services. That distinction matters where “AI crypto” often means a token with an AI label and little else. Base benefits directly: every x402 payment is a transaction on its network, and transaction volume is the fundamental metric that gives a chain value.

The trend also strengthens the stablecoin thesis. If agents pay in USDC by default, growth in agent commerce translates directly into stablecoin demand — not speculative demand, but transactional demand that persists regardless of market sentiment. That is the kind of demand that makes an asset class durable.

Investors should still keep perspective. Micropayment volumes, measured in dollars, remain small, and the autonomous share is smaller still. But the direction is clear: machines are becoming economic actors, and they are paying in crypto because nothing else works at their scale.

FAQ

How many transactions went through x402 on Base? More than 138 million over the 16 months ending in September, according to a Chainalysis study reported by PYMNTS.

How small are AI agent micropayments? More than 87% of payments were between a hundredth of a cent and 10 cents — amounts card processing cannot handle economically.

What is x402? An open payment protocol built on the HTTP 402 “Payment Required” status code: a server states its price, an agent pays in stablecoins such as USDC on Base, and the resource is delivered.

What do AI agents buy with crypto? Computing power, live web data, and coding tools — for example, fresh market data, a model response, or a code review, each paid for separately.

Is all x402 volume genuinely autonomous commerce? Not necessarily. TRM Labs found only 0.6% to 7.5% of x402 settlements looked truly agentic after filtering.

Source: PYMNTS

MM Crypto Analysis Web