The Agent Payment Market Settles in Cents

Two measurements landed this month that describe the same thing from opposite ends. One is a chain analyst counting every x402 payment on Base. The other is our own catalog, measured this morning. They agree on the shape of the market, and that agreement is more interesting than any headline number.

What the chain data says

Chainalysis published an x402 report on October 14, 2026. The cumulative count on Base reached roughly 138.5 million transactions by the end of Q3 2026, up from about 58,800 in Q3 2025. The payment-size distribution is the part worth reading twice: more than 87% of x402 payments were between one-hundredth of a cent and 10 cents, with the most common range between one and 10 cents.

The adoption curve inside that count is steep in a specific way. The median time from a wallet's first funding to its first x402 payment fell from 73 days in late 2025 to 18 minutes in Q3 2026. The share of wallets that received a second funding within 30 days rose from 70.6% for the Q3 2025 cohort to 89.3% for the Q2 2026 cohort.

Chainalysis also added the sentence that keeps the headline honest: transaction counts alone do not establish how much economic value the activity represents, and wallet refilling is not direct proof of commercial profitability. The report describes agentic commerce as still being in an experimentation phase.

What our price list says

We pulled /api/services this morning: 1,709 services, each with a price in whole cents. Sorted into buckets:

The median price is 1 cent. 90.0% of the catalog is priced at 10 cents or less. Twenty-seven services cost more than a dollar, and they are not the ones carrying the traffic.

The shape matches

The market's payments cluster between 0.01 and 10 cents. Our price list clusters between 0.01 and 10 cents. Whatever else is unresolved about agent payments, the unit of account has converged: this is a market that transacts in cents, and a catalog that wants to serve it has to price in cents too.

That convergence is a relief and a trap. It is a relief because it means the pricing question is settled for now — no elaborate metering, no subscription tiers, no per-seat negotiations. It is a trap because if everyone prices in the same band, price stops being a differentiator. The only remaining question is whether a call is worth making twice.

A cent is not a business model. A cent that gets charged again tomorrow is.

What our own usage says

Here is the uncomfortable half. Our live stats this morning report 44,379 lifetime trial calls across 5,442 unique users, and 282 externally settled calls with 7.11 USDC of gross settled volume. Refunds are reported separately: 98 refunds, 1.285 USDC. These two populations are not comparable — the trial side includes repeated probing and the platform's own verification traffic, and the paid side is gross settled value that is never adjusted downward when a refund happens. Do not divide them into a conversion rate. Read them as two different measurements of the same platform.

The per-service trial counts are more useful, because they have a denominator that matches the numerator. Summing trialCount across the 1,709 services gives 43,258 trial calls. Three services account for 25,987 of them — 60.1% of the total:

The remaining 1,706 services split the other 17,271 trial calls. The median service gets a number that rounds to zero. This is the long tail doing what long tails do: existing, being priced correctly, and not being called.

Why this matters for anyone selling an API to agents

Three practical conclusions, all drawn from the numbers above:

1. Supply is not the bottleneck. A catalog with 1,709 endpoints and a median of near-zero trials per endpoint has already answered the question "do we have enough services?" The answer is yes. Discovery and relevance are the bottleneck, which is a different problem with different work.
2. The example is the interface. An agent does not browse a landing page. It reads the 402 response, finds the input example, and either runs it or edits one field. If the example is ambiguous or contradicts the handler, the agent gets a confident answer to a question it did not ask, and never learns why. We audited 1,609 published input examples against their handlers; four could never satisfy the equality test in the handler that serves them, because the advertised value belonged to a different literal domain.
3. Measure repetition, not requests. A million requests from one retry loop is not adoption. The number that matters is how many distinct callers come back to the same endpoint on a different day. We have started treating repeat usage as the metric to defend, and treating raw request counts as something that bots and our own probes can inflate.

What we are changing

• Catalog descriptions are being rewritten in batches from a template into factual, service-specific text. The share of degenerate descriptions has moved from 21.4% to 15.6% in the first five batches of 100.
• The input-example audit is now a standing check, not a one-off: every example is compared against the handler's own comparison literals, and mismatches are queued for the next gateway build.
• Growth reporting now separates gross settled volume, refunds, and repeat callers. The platform's own verification traffic is excluded from customer-facing numbers rather than silently averaged in.

The market context, briefly

Cloudflare announced its Monetization Gateway on October 2, 2026, in closed beta: per-request, per-query, or per-token billing, settling in USDC on Base through Coinbase's x402 Facilitator, with sellers setting price rules by URL, headers, or query parameters. Four production use cases were named — AI inference, web search, market signals, and PDF generation — and availability is limited to eligible U.S.-based sellers and buyers for now.

That is a strong signal for the pricing model, and a reminder that the infrastructure layer is consolidating around HTTP 402. What it does not solve is the problem in this post: a payment rail makes a call payable. It does not make the call useful, discoverable, or worth repeating. That work stays with whoever publishes the endpoint.

Sources. Chainalysis x402 figures as reported by Brave New Coin, October 15, 2026: bravenewcoin.com. Cloudflare Monetization Gateway details as reported by Metaverse Post, October 2, 2026: mpost.io. First-party figures measured 2026-10-18 from /api/stats and /api/services; price buckets computed from the priceCents field, trial concentration from the trialCount field on 2026-10-18.