What 2,646 Paid Agent API Calls Reveal About the M2M Economy

Last week, Karim Gueye published an unusually honest breakdown of 2,646 settled x402 payments on his platform NetIntel — 194 distinct paying wallets across 101 endpoints. It's the most detailed public data we've seen on what AI agents actually pay for, as opposed to what they try for free.

We cross-referenced his findings with minia2a's own trial data (5,927 free trials, 316 unique agents, 228 services). The patterns converge. Here's what the combined data tells us about the M2M economy in August 2026.

1. Revenue is brutally concentrated

On NetIntel, 5 endpoints drive 69% of revenue. One text-to-structure endpoint alone captures 42%. Thirty-five of 101 endpoints had zero settled calls, ever.

On minia2a, the pattern is even starker: only 3 of 228 services have processed paid calls — CAPTCHA solving ($0.01 volume), gas estimation ($0.01), and knowledge recall ($0.01). The remaining 225 services are on free trials only.

This isn't a failure of either platform. It's the nature of an early M2M market: liquidity pools around a handful of high-utility endpoints, and the long tail is discovery inventory, not revenue inventory — yet.

2. Agents pay for text transformation, not exotic data

NetIntel's top earners: schema extraction, translation, structured inference, sentiment analysis. Their technically hardest endpoints — network intelligence, deep packet analysis — mostly never earned.

minia2a's top trial endpoints tell the same story from the demand side:

The lesson: agents pay to transform text and access infrastructure. They don't pay for exotic data. Build what agents need to function, not what sounds impressive.

3. The cheap-endpoint acquisition funnel

This is the most actionable finding from NetIntel's data:

"The cheapest utility endpoint was the first paid touch for 39% of all wallets. Those wallets later spent 3.4× across the catalog. 92% of wallets returned for more than one paid call."

Sub-cent endpoints aren't loss leaders — they're the on-ramp. Once an agent operator trusts the payment flow with a $0.001 call, they're willing to spend $0.05 on AI summarization or $0.10 on research. The first payment is a trust event, not a revenue event.

On minia2a, the $0.001 ping endpoint (payment flow test) has only 8 trials across 4 users — it's buried. Making ultra-cheap entry points prominent isn't just good UX; it's the primary conversion mechanism in agent payments.

4. Call volume ≠ revenue

NetIntel's two most-called endpoints combine for ~35% of call volume but only 6.3% of revenue. The third-most-called earns 42% of revenue alone. Frequency and value are decoupled.

On minia2a, Fear & Greed Index has the highest engagement (22.2 calls/user) but $0.01 total paid volume. Meanwhile, CAPTCHA solving at $0.001/call generated the same $0.01 volume from 8 paid calls. The high-frequency endpoints aren't necessarily the revenue drivers.

5. The market is real but early

NetIntel has deeper monetization. minia2a has broader discovery. Both are real — these aren't vanity metrics, they're on-chain USDC settlements.

The combined signal: ~230 wallets have made at least one agent-to-agent payment across two platforms. That's small in absolute terms but growing rapidly — x402 Foundation reported 75M monthly transactions in July, though most are from a few high-volume operators.

What this means for agent API builders

1. Ship a $0.001 utility endpoint first. It's not about the revenue — it's about getting the first payment through. That first paid call is the conversion event.
2. Text transformation > exotic data. Schema extraction, summarization, translation, sentiment — these are what agents pay for today. Build what's useful, not what's impressive.
3. Expect a power-law revenue distribution. 5% of your endpoints will drive 70%+ of revenue. That's not a bug — it's the market finding product-market fit. Your job is to make sure the other 95% are discoverable for when agents need them.
4. List on multiple marketplaces. NetIntel and minia2a serve different discovery surfaces. The agents finding you on one platform may not overlap with the other.

The discovery layer is still up for grabs

Neither NetIntel nor minia2a has solved discovery. Both have supply (100+ endpoints). Both have some demand (dozens to hundreds of wallets). But no platform has cracked the distribution problem — how does an agent building a new workflow find the right x402 endpoint in seconds?

That's the opportunity. The payment rails work. The standards exist (x402 Foundation, 40 members). The marketplaces are live. What's missing is the search-and-match layer that connects an agent's intent ("I need to solve a CAPTCHA on this page") to the right paid endpoint, with prices, reliability scores, and one-click payment.

That's what we're building at minia2a. 228 services, 5,927 trials of training data, and growing.