The x402 Bazaar lists thousands of pay-per-call APIs for AI agents. Most of them never get paid. We watched every listing, every day, for a month and trained a model that tells you which ones will.
Scroll for the story ↓
The Coinbase CDP discovery index, the "Bazaar", is where agents find x402 services to pay. On any given day it lists 14,000 to 16,000 of them, from about 2,000 hosts. Some hosts publish hundreds of near-identical endpoints; one lists 840.
Each dot in the panel stands for a slice of that index.
In a typical week, only 24% of listed services receive even one paid call. Over two weeks it is 32%.
If you are an agent choosing an endpoint, or a builder deciding what to list, you are working against that base rate.
Of the listings present on 24 August 2026, 59% were gone thirty days later. In the same month, as many new ones appeared.
A listing is a promise, and most promises lapse.
The index removes a service exactly 30 days after its last paid call. So whether a listing survives is mostly arithmetic on a timestamp you can already read.
We ask the question that matters instead: will anyone actually pay for it in the next 7 days?
From 24 August to 23 September 2026 we captured the whole index once a day: 31 snapshots and 330,035 labelled listing-days.
For each one we know whether a paid call followed within 7 and 14 days. That record is ours, and it is not published.
We fine-tuned Laya, a 421M-parameter decision model: a ModernBERT encoder with a calibrated decision head. It reads the URL, the description, the price and the network.
Out comes a probability, not a paragraph. There is nothing to parse and nothing to hallucinate. It runs on an ordinary CPU, no GPU, in about two seconds.
A model that has seen a host's other endpoints can just memorise the host. So we tested only on later dates and on 442 hosts the model never saw: 13,644 listing-days in all.
Every number on this page comes from that held-out test.
Reading the listing alone, the free model ranks services at AUC 0.64. That's weaker than a model that can also see call history. But on brand-new listings, where there is no history, it's the only model that beats chance.
Add the Bazaar's own telemetry and it reaches 0.73, beating the statistical baselines (0.67–0.68). The top tenth of its scores got paid 65% of the time; the bottom tenth, 3%.
Paste an x402 listing as it appears in the Bazaar, or one you haven't published yet. The free model reads only what's below.
Your probability appears here. For reference: 24% of listed services get a paid call in a given week.
The model behind this demo, open-weights on Hugging Face under Apache-2.0. It loads with laya.load() and runs on a CPU.
The stronger model, scored against today's Bazaar data. It's sold per call, paid in USDC over x402, so an agent can check an endpoint before it pays one.
What it is not. It predicts demand, not quality or honesty. A scripted self-call counts as a paid call. It's trained on one index over one month, in a market that is weeks old, and it's retrained monthly. The free model is over-confident above about 50%: treat 70% as roughly 45%.
