More than 138 million payments using the x402 protocol had occurred on Base by the end of Chainalysis’s 2026 reporting period, according to a North America study published October 14. The count shows substantial use of a rail designed for software to buy online services in small increments. It does not, by itself, show that 138 million purchases were independently chosen by artificial-intelligence agents.

Small payments, large count

Chainalysis examined x402 activity over an extended 16-month period ending in September 2026 because the technology is new. More than 87% of the payments it measured were between $0.0001 and $0.10, with one to ten cents the most common range. These are transaction counts, not merchant revenue: the firm did not provide a total dollar value for the x402 payments in this section of its report.

Related reporting: Block brings Bitcoin Lightning payments to x402

x402 adapts the web’s HTTP 402 Payment Required response into a programmatic checkout. Software asks for a resource, receives a price and returns a payment authorization; a facilitator can then verify and broadcast the on-chain settlement. That structure can make metered access to data, code tools or model inference practical without asking a person to complete a card checkout for every request. Chainalysis identified spending at information services, developer tools and AI-model providers, alongside less clearly productive categories.

New wallets appear to spend sooner

The report also tracks how quickly funded wallets begin paying. In late 2025, the median wallet was 73 days old when it made its first x402 payment; by the third quarter of 2026, the median interval had fallen to 18 minutes. The share making a first payment within an hour of funding rose from 4% in the third quarter of 2025 to 67% a year later. Chainalysis interprets the shift as evidence that more wallets are being set up for a specific payment task. The timing alone cannot establish whether the payer is an AI system, a conventional script or a person using automation.

Among transactions Chainalysis could geolocate, U.S.-linked wallets accounted for 88.1% of x402 payments for AI-model inference and 41.8% for agent frameworks and tool catalogs. Those percentages apply to identified transactions in specific categories, not all x402 activity or every country’s software economy. Geographic attribution and categorization are analytical estimates rather than a direct census of users.

Why the agent label needs caution

A separate September analysis by TRM Labs illustrates the identification problem. It found about 198.9 million x402 settlements across Base, Solana and Polygon since May 2025, then applied filters intended to remove self-payments and other flows unlikely to represent ordinary commerce. Of the remaining value, TRM estimated that only 0.6% to 7.5% appeared agent-driven under its two models. Its chain coverage, window and method differ from Chainalysis’s Base count, so the figures should not be subtracted from or directly compared with one another.

TRM cautioned that a single-purpose agent repeatedly buying one fixed-price service can resemble a script, while self-dealing or testing can inflate raw protocol activity. Chainalysis likewise describes agentic commerce as experimental, even while pointing to faster wallet funding and recurring top-ups. The evidence supports growth in machine-readable payment infrastructure; the scale of genuine autonomous buying remains uncertain.

What the results mean

For developers selling API access or model calls, the report suggests a reason to test pay-per-use prices smaller than card networks can comfortably handle. For compliance teams, many tiny, rapid settlements shift attention toward who funds a wallet, what budget a program may spend and which services receive payments. Neither report demonstrates that x402 will displace traditional checkout or that transaction growth will translate into durable merchant income. Better attribution of real customers, software behavior and settled value is the next test.

Sources

AI-generated editorial image; not a photograph of the reported event. Prepared with AI assistance and source verification.