What machine commerce claims — and what actually happens

With x402, software can pay for an individual web request. A service quotes a price for a particular web address, called a route. The site where the route runs is its host. The buyer’s software signs a payment in a stablecoin, a token designed to hold a steady value. A payment service called a facilitator submits it to the blockchain. Public registers such as x402-trust and x402-list group payments by facilitator.

x402scan reported $1,654,885 for the 30 days ending 5 November. x402-trust says it uses x402scan’s list of facilitator addresses to build its own, so those two registers share a source.

FluxA accounted for most of the reported payment value at month-end. On 4 October, x402-trust reported $935,288 in payments for the preceding 30 days. By 31 October its 30-day total had risen to $1,489,633. Over the same period, the amount it reported for FluxA rose by $659,672, more than the increase across all facilitators combined. For the period ending 31 October, x402-trust reported $522,448 for Coinbase. Coinbase handled 35.1% of the reported payment value and 67.6% of the settlements.

We checked every incoming transfer of USDC, a dollar stablecoin, at the 2,273 addresses for all 31 days of October. We counted transfers made under a signed authorisation. Two independent public copies of the Base blockchain gave the same results.

The record shows 1,501,232 payments from 11,517 wallets. The dollar total was $50,331 after excluding $32 in our own payments.

The filters start with the $50,331 traced to the listed addresses. Each successive filter applies a stricter rule for identifying a customer purchase. The final row shows what those rules would imply under two assumptions about the wider market.

The published testers are five services that name their paying wallets: vet402, nohumans.directory, Agent402.tools, AgentPay and Sippar.

The address quoted by ax1’s token-grading routes received $13,531 of the central estimate: 676,616 payments of about two cents, from 6,640 wallets. Without it the central estimate is $18,761.

At 150 addresses, 591 wallets paid an address that had already sent them USDC. At the address quoted by api-xpay’s route, $10,288 went out to 277 wallets and $10,286 came back from them, in 66,129 payments.

A listing in Coinbase’s Bazaar catalogue lapses after 30 days without a settlement. One seller told us it buys from its own routes and identified the wallet it uses. That wallet paid the seller’s address $26 in October.

• We can miss a seller paying from another wallet, or a tester whose wallet is unpublished, so the true figure can be lower. We can also exclude real customers who received a refund, bought from many addresses or bought just once, so it can be higher. A repeat purchase at the listed price still leaves the buyer’s identity uncertain.
• The measured figures cover the Base addresses quoted to us. Sellers in no catalogue, payments on Solana or by card, and payments to an address other than the one quoted are outside them and would need separate measurement.

• When a route quotes a gateway’s payment address, we can see what reached the gateway. How much the supplier ultimately received is unknown to us.
• A signed authorisation does not prove a purchase. Two listed hosts quoted the zero address, which is no one’s wallet, as their payment address. $11.1 million went there in transactions that also used an authorisation, all from one address and none submitted by a listed facilitator. It is not a payment to a seller, and we set it aside.

A further $897,220 reached the same addresses as ordinary transfers, with no authorisation. We do not count it.

Coinbase’s Bazaar, a catalogue of x402 services, lists a route after Coinbase’s facilitator settles a payment for it. Of the 500 largest payments we saw at the quoted addresses, each worth $10 or more, 488 were submitted by Coinbase facilitator addresses.

Using figures reported by x402-trust, we estimate buyer payments of $202,000 to $1,006,000 for the 30 days ending 31 October. At the low end, the strictest filter retains 39% of the dollars at listed addresses. Applying that share to the $522,448 reported for Coinbase gives the lower figure. At the high end, the broadest filter retains 68%. Applying that share to the full $1,489,633 reported by x402-trust gives the upper figure. The transfers traced to listed addresses amounted to less than a tenth of the Coinbase figure. The buyer share elsewhere could be higher or lower, leaving too little evidence for a central estimate.

For the Machine Payments Protocol (MPP), a second way for software to pay for a web request, MPPscan reported about $7,400 across 295 servers for the 30 days to 5 November. How much MPP buyers paid sellers remains an open question.

Blockchain data and on-chain tools accounted for 7.2% of the x402 hosts that quoted a price and 48.5% of the October dollars we estimate buyers paid at the addresses we read. One seller, at the ax1 address in Finding 1, accounted for most of that category’s dollars. AI model services accounted for 4.3% of those hosts and 17.9% of estimated buyer payments. Another 15.9% reached addresses shared by hosts in different sectors, so we could not assign those payments to a sector.

We classified the 2,152 hosts from which we received a price quote in October. The quote arrives in a reply called a payment challenge, which shows that a price was offered, not that anything was paid. Two AI models, working independently, each labelled every host against one written rulebook, from the seller’s own description of its routes and never from the host’s name. They agreed on 1,935 hosts (89.9%); that measures agreement, not accuracy. A host they disagreed on is shown as unresolved.

By kind of service, 1,283 of the 2,118 x402 hosts (60.6%) sell a data lookup and 98 (4.6%) run a model.

The median listed price was USD 0.01 across the 15,353 endpoints that quoted a USDC price in October. Of those endpoints, 41% quoted less. Among the 11,766 endpoints priced in both September and October, 97.3% ended October at the same price as in September.

The host shares are not revenue or usage. The dollar shares cover the sellers we read. Our list of routes was last refreshed in mid-October, so hosts listed after that are not in it. Descriptions are the sellers’ own words.

x402-trust reported the figures below for the 30 days to 31 October. The shares and averages are our arithmetic over its rows. By those figures, FluxA led by dollars and Coinbase by number of settlements.

• The dollar leader changed during the month. In the 4 October reading Coinbase had 54.4% of the register’s dollars and FluxA 19.8%. FluxA has led in every reading since 21 October.
• x402-trust’s own all-time totals fell for 20 of its 32 facilitators between our readings of 2 and 3 October, and for 14 the next day. The reason for the change is unclear. None fell in the 25 later pairs of readings.
• A second register, x402-list, tracked 40 facilitators and reported $2,332,901 for the same 30 days. Twelve names on x402-list did not appear on x402-trust’s list for the same period.

Of the $1,906, $1,827 came from one wallet, and the receiving address had sent that wallet $1,914. The last column estimates how much of the October total came from customers, using the central filter in Finding 1.

Each week in October we tested routes from a fixed panel of 755 listed x402 routes priced at five cents or less, with at most two per host. We used USDC on Base and built requests from the inputs each route published in its payment terms. The results describe this low-price panel, which may not be representative of the wider market.

Of 752 routes called on 7 October, 564 settled the payment and returned an answer. Of the 755 called in each of the next three weeks, 542, 527 and 512 did. Over the same period, calls that sent no payment rose from 65 on 7 October to 125 on 28 October. An answer here means a successful reply with a readable body; we did not judge its content.

The inputs matter. Of 394 routes called on 30 September with no inputs and on 7 October with their published inputs, 195 answered on 30 September and 299 on 7 October. In the earlier comparison, using the same request on 23 and 30 September, four routes changed outcome in each direction.

On 54 calls, payment went through and we received no successful reply with a readable body. Those calls spanned 22 routes on 14 hosts. Of the $0.84 paid, $0.08 was returned.

• On 18 calls the route rejected its own published placeholder (the word example, for instance) after the payment went through.
• We sent complete requests on 12 calls across 5 routes and 4 hosts.
• On 9 calls our request left out an input the route declared.
• On 7 calls the error named an input the payment terms did not declare.
• On 6 calls the route turned away its own published example, which was not a placeholder.
• On 2 calls the limit was our tool’s: it asks the price in one request and pays in a second.

On agentworld, two routes asked for payment again after the transfer had gone through, in all four weeks. Two routes, one on Sponge and one on Orthogonal, sent the money back each time, which is the $0.08 above.

Each week, 86 to 103 calls returned an error with no transfer we could find. On 11 to 19 calls a transfer went through and we received an error or no answer. Across 76 calls on 25 routes on 16 hosts, the error message said the caller had incurred no charge. We found no transfer on 74 of those calls. The other two readings were ambiguous.

All 2,175 transactions that a reply named were on the blockchain for the quoted amount to the quoted address. On 84 other paid calls we could read no reference (see Other signals).

Some routes rejected the example in their own payment instructions. 26 routes publish a placeholder where the protocol asks for an example value: the word example, a type name such as string, or the input’s own name. A route can fall in more than one group.

• 16 of them, on 12 hosts, turned it away at no charge (60 calls). Counting by hand the placeholders our written rule misses, this group is 33 routes on 20 hosts.
• 5, on 3 hosts, answered with an error after the payment had settled: the 18 calls above, $0.41.
• 5, on 4 hosts, accepted it and returned an answer (17 calls).

On six hosts, five of them sharing one domain, the error message says in words that the value sent was a placeholder.