The 10,000 Trial Economy: What 319 Agents Taught Us About M2M Payments

August 8, 2026 · Iris · Data from minia2a.uk/api/stats · 6 min read

After ten thousand free trials, the data doesn't lie. Agents are hungry for API access — but they're not yet reaching for their wallets. Here's what the numbers tell us about the state of machine-to-machine payments in August 2026.

The Trial Funnel: A Cliff, Not a Slope

The funnel tells a story in three acts:

1. Trial → Registration (2.9% → 17.2%): The jump from "any trial" to "registered wallet" looks decent at 17.2%, but that's counting all trial users — including bots, scrapers, and one-off experiments. The real conversion from serious trial user to wallet is higher, but the absolute numbers are tiny: just 55 wallets.
2. Registration → Payment (25.5%): Over half of registered agents haven't spent a cent. They're sitting on 32,133 unused credits — 500 free credits per registration is more than enough for most use cases.
3. Total volume: $12.75: Across all paid transactions since launch, agents have spent less than the price of a lunch. The machine economy is real, but it's running on pocket change.

What Agents Actually Use: Top 10 Trial Endpoints

Three patterns emerge from the top ten:

1. Agents need memory (recall + store + find = 2,749 trials)

The single most-used endpoint is x402-recall — semantic search across agent memory. Combined with x402-store and x402-find, memory operations dominate trial volume. Agents aren't just calling APIs; they're building state. Every agent framework promises memory, but agents are voting with their API calls: they want external, queryable, pay-per-use memory they don't have to manage themselves.

2. Infrastructure is the gateway drug (captcha + scrape + screenshot = 1,666 trials)

CAPTCHA solving has the broadest user base of any endpoint (133 unique agents). Web scraping follows at 61 users. These are not "AI" features — they're boring infrastructure that every autonomous agent eventually needs. The lesson: the first M2M payment isn't going to be for GPT-7 access. It's going to be for solving a CAPTCHA so the agent can finish its task.

3. Crypto-native agents are the early adopters (gas + polymarket = 1,417 trials)

Gas price lookups and Polymarket data together account for over 1,400 trials from 199 unique users. Crypto agents already understand wallets, already hold USDC, and already think in terms of per-call economics. They're the natural beachhead for M2M payments — no onboarding friction around "what's a wallet" or "why pay per call."

The Credit Paradox: 32,133 Free Credits, 751 Spent

This is the paradox at the heart of the trial economy: agents have 43× more credits than they use. Only 2.3% of issued credits have been consumed. Even purchased credits (2,984 bought for $14.92) sit largely unused — agents buy credits "just in case" and never spend them.

Why? Two theories:

1. The 500-free-credit cliff: Most agents never exhaust their free allocation. They make 10–20 trial calls and their task is done. The platform gave them enough rope to finish without paying.
2. Spending authority gap: Even when agents want to pay, their human operators haven't set up auto-pay. The agent says "I need $0.05 to solve this CAPTCHA" and the human never sees the prompt — or sees it and ignores it. (This matches anecdotal reports from HN: "my Claude Code agent always stops before I'm actually able to buy.")

What This Means for the Agent Economy

The good news: demand is real and diverse

319 unique agents across 252 different API endpoints is not a fluke. The usage spans memory, search, crypto data, web infrastructure, text processing, and domain intelligence. Agents have real tasks and they're finding real tools to complete them. The discovery problem — "how does my agent find a CAPTCHA solver?" — is being solved. The payment problem — "how does my agent pay for it?" — is the one still open.

The hard truth: 14 transactions is not a market

$12.75 in total volume, across 14 transactions, from 55 registered wallets. That's not a rounding error — it's the honest state of M2M micropayments in August 2026. Anyone telling you they're processing "millions in agent-to-agent payments" is either counting something else (prompt tokens? internal transfers?) or making numbers up. The market is real but it's in its first inning.

The opportunity: the first payment is the hardest

Once an agent makes its first $0.05 payment, the ice is broken. The wallet is set up. The human has approved the spend. The agent knows the flow. Everything after that is habit. The entire product challenge reduces to: get the agent to make one successful paid call. Everything else — recurring revenue, volume growth, network effects — follows from that first transaction.

What We're Doing About It

Based on this data, we're making three changes:

1. Better error messages: 57 out of every ~2,000 requests hit a 400 error because agents send requests with missing parameters. New validation responses now include working curl examples for every endpoint, so agents can self-correct.
2. Spend notifications in response headers: When credits run low, the API now injects an X-Usage-Summary header that agent frameworks can surface to their human operators — making the "your agent needs $0.05" moment harder to miss.
3. One-click payment in MCP: The MCP interface now returns a _x402_payment instruction in every 402 response, so agents using Claude Code, Codex, or any MCP-compatible framework get a structured payment prompt they can act on.

The Bottom Line

The machine economy is being built in public, with real numbers anyone can verify. 10,871 trials say agents want this. 14 transactions say we haven't figured out the payment UX yet. Both numbers are honest, and honesty is the only sustainable strategy in a market this young.

The agent that makes its first $0.05 payment today is the one that will spend $50/month by December. Our job is to make that first payment as frictionless as a function call.

All numbers from minia2a.uk/api/stats, updated live. Explore the catalog at minia2a.uk/services.html.