Key takeaways
- x402 is a Coinbase standard that lets software pay for online resources through a web request, without a human approving each purchase. It runs on Base.
- Chainalysis data shows 89.3% of wallets first funded in Q2 2026 received another deposit within 30 days, up from 70.6% a year earlier — and the median time from funding to first payment fell from 73 days to 18 minutes.
- An academic study of 136.7 million settlements found 21.2% were fictitious and 63.78% happened inside clusters of wallets linked by funding relationships.
- Of 25,163 listed resources on Base, 811 distinct recipients were identified — and only 249 earned at least $10 during the observation window.
- The settlement count is the one metric an interested party can manufacture almost for free, because the facilitator sponsors the gas and nothing onchain records who controls a wallet.
Here is a number that has been travelling fast this week: x402, Coinbase's machine-payment standard, has processed more than 138 million transactions on Base. Alongside it, Chainalysis data published this week shows that 89.3% of wallets first funded in the second quarter of 2026 came back for another deposit within thirty days.
Those are real figures from a credible firm, and the retention one is genuinely striking. A year earlier the same measure was 70.6%.
But three separate teams have now gone looking at what sits underneath the settlement count, using three different methods. They disagree about plenty. They agree about the thing that matters.
What x402 actually does
Start with the mechanism, because the whole argument depends on it.
x402 revives an HTTP status code that has sat mostly unused since the web was designed: 402 Payment Required. A piece of software requests a resource — an API call, a model inference, a dataset — and instead of a login prompt it receives a payment instruction. It settles in stablecoin, and access is granted. No human clicks anything.
That is a real capability and a genuinely interesting one. It means a program can buy things in the course of doing its job, in amounts too small for a card network to bother with. More than 87% of x402 payments over the sixteen months to September fell between $0.0001 and $0.10, with one to ten cents the most common size.
The behavioural numbers are the strongest part of the case. Two-thirds of newly funded wallets made a payment within an hour, up from 4% a year earlier. The median gap between funding a wallet and its first payment collapsed from 73 days to 18 minutes. Whatever is happening, it is not people carefully deliberating.
Which is exactly the problem.
The measurement nobody wanted to do
In July, researchers from City University of Hong Kong, the Chinese University of Hong Kong and Zhejiang University measured every x402 settlement on Base across a 280-day window — about 136.7 million settlements worth roughly $44.1 million.
Then they asked who was on each end.
- 21.2% were fictitious — self-payments and closed loops, a wallet paying itself
- 63.78% were internal settlements inside a cluster of wallets linked by funding relationships
That second figure deserves care, and the authors give it care. Wallets funded from the same source are not automatically fake. A business legitimately runs many wallets. But it is not independent commerce either — it is one party moving value around its own perimeter, and counting it as adoption is a choice rather than a finding.
What they could bound was the genuinely independent portion: somewhere between $187,861 that demonstrably reached a nameable service with no detectable link to its payer, and $20.26 million — 45.92% of value — that simply could not be proven manufactured.
The gap between those two numbers is the honest state of knowledge. Anyone quoting a single confident figure for agentic commerce is quoting the end of that range they prefer.
Then two more teams found the same shape
TRM Labs ran a different method across a wider base — roughly $52.7 million across 198.9 million settlements on Base, Solana and Polygon. After excluding self-payments and other anomalous flows, it estimated that between 0.6% and 7.5% of the remaining commerce by value looked like it came from AI agents at all.
Copper's research note took a third angle and found that a single launch month of Base wallets paying themselves accounts for 44% of the total. Strip the manufactured flows and genuine settlement on Base lands near $29 million.
These three estimates are not comparable to each other, and it would be sloppy to average them. The academic paper reports shares of settlements; TRM reports an agent share of screened commerce by value; Copper reports a Base total after removing self-payment flows. Different denominators, different questions.
What makes them worth reading together is not the numbers. It is that three teams with different data, different definitions and different incentives all pulled the same thread and all found the headline count detached from independent economic activity underneath it.
Why the count is the wrong metric, structurally
This is the part I find genuinely clarifying, and it is the paper's real contribution.
In x402, the facilitator sponsors the gas. A settlement therefore costs its originator approximately nothing. And nothing onchain records who controls a wallet — that has to be inferred from funding patterns afterwards, which is precisely what these studies spend their effort doing.
Put those two facts together and the settlement count becomes the one metric an interested party can manufacture almost for free. Not through fraud, necessarily. A team testing its own integration at scale produces the same onchain signature as a team inflating a number, and neither has to intend anything improper for the total to stop meaning what people read it as meaning.
This is not a new failure mode. It is the same structure as unique-address counts in 2017 and total-value-locked in 2021 — a number that is cheap to move and expensive to verify, which is a combination that reliably ends up quoted in decks.
The number I would actually watch
Of 25,163 listed resources on Base, the researchers identified 811 distinct recipients. Of those, 249 earned at least $10 during the observation period.
Two hundred and forty-nine.
That is the figure that tells you where agentic commerce is, because unlike settlement counts it is expensive to fake: it requires somebody outside your own cluster to send you money for something they wanted. It is a small number. It is also a real one, and it can grow.
Set against it, the geographic split Chainalysis found is more interesting than the totals. US-attributed wallets account for roughly 30% of identified agent-to-service transactions but 88.1% of AI model inference payments. China-linked wallets account for about 90% of payments for security and code-analysis services, and 0.1% of model inference — which Chainalysis suggests reflects reliance on open-weight models such as DeepSeek and Qwen rather than paid API calls.
That is a specific, structural claim about who pays for what, and it is harder to manufacture than a transaction total.
What this does and does not mean
It does not mean x402 is fake. The standard works, the plumbing is real, and a mechanism for software to pay sub-cent amounts without a human in the loop is a legitimately new thing. BlackRock's digital assets research team is reportedly watching the sector for exactly that reason — continuous small-payment demand would be a genuinely different kind of flow.
It does mean that the specific number being used to argue adoption does not measure adoption, and that three groups have now shown this with evidence rather than assertion.
The practical habit, and it generalises well beyond this story: when a metric is cheap for an interested party to move, ask who benefits from it being large before asking whether it is large. We looked at a related case recently when an AI model found a decade-old bug nobody else had — there the striking claim held up under checking. Here the striking claim is a count, and counts are the cheapest thing in this industry to produce.
Watch the 249. It is a much less impressive number, and it is the one that would actually mean something if it moved.
Frequently asked questions
What is x402 in plain terms?
It is a standard from Coinbase that uses the long-dormant HTTP 402 'Payment Required' response. Software requests a resource, gets a payment instruction instead of a login prompt, settles in stablecoin, and is granted access — with no human approving the purchase. It runs on Base.
Does this mean the 138 million transaction figure is false?
No. The settlements happened. The argument is about what they measure. An academic study found 21.2% were self-payments or closed loops and 63.78% occurred within clusters of wallets linked by funding, so the count reflects activity inside a small number of perimeters rather than independent commerce between unrelated parties.
Why can't anyone say exactly how much real agent commerce there is?
Because nothing onchain records who controls a wallet — it has to be inferred from funding patterns afterwards. The study bounds it rather than pinning it: from $187,861 that demonstrably reached a nameable service with no detectable link to its payer, up to $20.26 million that simply could not be proven manufactured. Anyone quoting one confident number is picking an end of that range.
Is the 89% refill rate meaningful on its own?
It is a real measurement and the direction is notable — 89.3% for wallets first funded in Q2 2026 against 70.6% a year earlier. But the denominator only includes wallets that transacted on at least two separate days, single-day wallets were excluded, and cohort sizes were not disclosed. It tells you funded wallets keep being funded; it does not tell you who is funding them.
Has this pattern happened before in crypto?
The structure is familiar. Unique-address counts in 2017 and total-value-locked in 2021 shared the same property: cheap for an interested party to move, expensive for an outsider to verify. That combination tends to produce a widely quoted figure that drifts away from what people think it measures.
What should I watch instead of the transaction count?
The number of recipients earning real money from parties outside their own funding cluster. In the study that was 249 recipients earning at least $10, from 25,163 listed resources. It is unimpressive and hard to fake, which is what makes it informative.
Sources
Every figure and quote above traces to one of these. Checked on publication day.
- cryptoslate.com — AI agents keep coming back for crypto as x402 wallet refill rate surges to 89%
- arxiv.org — How Agentic Is Agentic Commerce? A Population-Scale Measurement of x402 Adoption and Authenticity (arXiv:2607.12575)
- trmlabs.com — Who's actually paying? Measuring AI agent payments onchain — TRM Labs
- decrypt.co — AI Agents Spending Money Online? New Research Says Not Really — Decrypt
- copper.co — Opening Bell issue 65 — Copper research
Deepmoni Das, Editor
I read the primary filings, exchange notices and post-mortems behind each story and write up what they actually say — with every source linked so you can check me. I am not a trader and I do not hold positions in anything covered here. Research and drafting are AI-assisted; the sourcing, the edits and the mistakes are mine.
Not financial advice. Crypto assets are volatile and unregulated in many jurisdictions. In India, gains are taxed at 30% with 1% TDS on transfers. Do your own research and never invest money you cannot afford to lose.
Editorial note: Crypto Shakti uses an AI-assisted research and drafting workflow. Every article is grounded in the linked primary sources and live market data captured at publication time.