What AI agents actually pay for — six weeks of data from 101 pay-per-call endpoints

Six weeks of production data from NetIntel, a platform for pay-per-call APIs settled in USDC, reveals that five endpoints drive 69% of all revenue, with a single text-to-structure endpoint accounting for 42% alone. The data shows that agents pay for transformation—schema extraction, structured inference—not raw data access, and that spending tracks task value, not call frequency.

A few weeks ago I wrote up what agents were paying for on NetIntel https://netintel.dev , my platform of pay-per-call APIs settled in USDC over x402 https://www.x402.org/ — no signup, no API keys, no accounts. An agent hits an endpoint, gets a 402 Payment Required , pays a fraction of a cent, and gets structured data back. That's the whole loop. Since then the dataset has grown, I've instrumented every settled call into a proper database payer wallet, endpoint, price, latency, transaction hash , and I've launched a second settlement rail. So this is the rewrite with real numbers instead of eyeballed ones — and the findings didn't soften. They sharpened. 2,646 settled paid calls from 194 distinct paying wallets, across 101 live endpoints , over six weeks of instrumented production data. Every call in this dataset is a real on-chain payment with a transaction hash — no test traffic, no estimates. Settlement runs on Base, and as of this month on Solana too. This is still one platform's data in a young ecosystem — the caveats are at the bottom, and one of them is bigger than it looks. But the shape has now held for six weeks straight, and it's the same shape I flagged the first time. Five endpoints drive 69% of all revenue. One of them — a text-to-structure endpoint that takes messy input and returns strict typed JSON — is 42% by itself . The rest of the top five are all in the same family: translation, structured LLM inference, and one domain-intelligence report. The other 96 endpoints split the remaining 31%. Thirty-five of the 101 have never been paid for once. Not "underperformed" — zero settled calls, ever. When I first published this pattern I wondered if it was an artifact of a small sample. The dataset has since more than doubled and the concentration ratio barely moved. I now treat it as the market talking, not noise. Here's the part I'd want to know if I were reading this: that 42% endpoint is essentially one buyer — a wallet running a daily cron that has settled a payment nearly every day for six weeks. Mid-month I doubled that endpoint's price. The buyer didn't pause for a single day. When an agent has wired your endpoint into a workflow that works, the price of the call is nothing next to the cost of replacing it. Which leads directly to — I built NetIntel as a network intelligence platform — DNS forensics, IP reputation, certificate transparency, the works. What actually earns is transformation : schema extraction, translation, structured inference, sentiment. Text goes in messy, structure comes out guaranteed. Four of the top five earners are transformation endpoints. The network-intelligence catalog I was proudest of — the technically hardest stuff — makes up most of the 35 endpoints nobody has ever paid for. The reframe from the first article still holds, so I'll restate it plainly: an agent's expensive problem isn't accessing data, it's trusting the shape of what it already has. A blob of text it must parse itself costs tokens, a fragile parsing step, and an entire error branch for "what if this comes back malformed." A guaranteed clean structure removes that liability in one call. The most valuable thing I sell isn't information — it's the removal of the agent's own uncertainty. This one got more extreme with more data. The two most-called endpoints on the platform — a cheap utility lookup and a content extractor, together nearly 35% of all call volume — combine for 6.3% of revenue . The third -most-called endpoint, on nearly identical call volume, is the text-to-structure endpoint making 42% . Spend tracks the value of the task completed, not the frequency of the call. If you optimize for the metric that's easiest to see on a dashboard — call count — you will double down on exactly the wrong endpoints. This is the one the new instrumentation surfaced, and it complicates the "kill the breadth" conclusion from my first write-up. I track which endpoint each wallet paid for first . The winner isn't close: my cheapest utility endpoint was the first paid touch for 39% of all wallets. And those wallets went on to spend, across the whole catalog, 3.4× what that endpoint itself has ever earned . The pattern repeats down the acquisition table: cheap, low-stakes, easy-to-verify calls are how agents try a new provider. Then the cross-sell data shows them fanning out — more than half of all wallets have paid for at least two different endpoints, and 92% of wallets came back for more than one paid call . So the catalog isn't just dead weight plus five winners. It's a funnel: The 35 never-called endpoints are still dead weight. But I'm no longer treating "low revenue" alone as a kill signal — the question for each endpoint is now does it earn, or does it acquire? An endpoint that does neither, goes. None of that changes the headline, which has now survived a doubling of the dataset: agent spend is extreme in its concentration, it favors transformation over raw data, it ignores call volume entirely — and the cheap calls that earn nothing are how the expensive relationships start. If you're building for this market, plan for a few things carrying everything, figure out which few as fast as you can, and don't delete the front door while you're trimming. I run NetIntel — pay-per-call network and structured-data intelligence for agents over x402, on Base and Solana. If you're building agents and want to compare notes on what your traffic actually pays for, I'm genuinely interested; that's the data I wish more people published.