I benchmarked 41 pay-per-call x402 APIs: here's what actually works

Most benchmarks in the x402 / AI-agent space are marketing in disguise: one vendor's chart, zero reproducible method, no raw numbers. So when we shipped our fleet of pay-per-call APIs on Base, we decided to benchmark them the way we'd want to read it — same queries, same time window, published method, and every number either real or explicitly marked as missing.

This is the first post in the series. It covers the method and our own baseline: real latency and cost numbers from 7 live endpoints, measured today.

1. The problem agents actually have

An agent that can't reach the live web is a very expensive parrot. It hallucinates prices, invents news, and cites sources that don't exist. The fixes are obvious — plug in a search API, a price API, a market-data API — but every option comes with friction:

• API keys and accounts for each provider. Five services = five signups, five bills, five ways to leak a key from an agent's context.
• Subscriptions that punish experimentation. $50/mo to test whether search quality is even good enough for your use case.
• Zero verifiability. You get billed; you trust the invoice.

x402 flips this: the agent pays per call in USDC, on-chain, no account. A machine paying a machine. That's the thesis we built on — 41 pay-per-call endpoints covering web search, AI verdicts, token prices, market data, contract audits, AML screening, document extraction, trend intel, sentiment, GPU brokerage, and more. No signup, no subscription, no key management.

But "x402 makes payment easy" is a claim, not a proof. What agents actually care about is: does the data work, and is the price fair? So we benchmarked.

2. The method (steal this)

We picked the hardest fair fight we could find: web search quality. Search is the backbone of every research agent, and there are credible incumbents — Exa and Tavily — with published pricing. Here's the protocol:

1. One fixed query set — 25 queries across 5 groups: crypto prices (real-time), tech news (last 24h), niche Vietnamese market info, long-tail factual questions, adversarial/ambiguous queries designed to break naive search.
2. Same time window — all providers queried within a single 30-minute window to control for index freshness.
3. Identical downstream consumer — the same agent prompt and the same answer-extraction step, so we're measuring the search layer, not prompt luck.
4. Scored blind — 3 reviewers score 0–2 on factuality, coverage, and citation quality without knowing which provider produced which answer.
5. Cost logged per query — what each provider actually charged for the exact call, not the marketing "starting at" price.

Our websearch endpoint runs on Serper under the hood — we're not hiding that. The question isn't "who has secret tech," it's "does a $0.010 x402 call compete with the incumbents on quality, and what does the math look like at agent scale?"

Status of the head-to-head: the Exa/Tavily comparison ships as Part 2. What follows is Part 1: our own baseline, measured live today — real latency, real cost, real output samples from 7 endpoints. No cherry-picking; these are the first queries we ran.

3. Baseline measurements: 7 endpoints, real numbers (2026-10-19)

All measurements taken 2026-10-19 ~11:14 UTC via free-trial calls (no payment) from a single client. Latency = wall-clock round trip including network.

Endpoint	Price/call	Median latency	What we queried	What came back
news	$0.002	2.2s	"bitcoin price"	20 items, real headlines w/ timestamps + sources (CoinMarketCap, Robinhood, Investing.com, Yahoo Finance)
fxrates	$0.004	1.7s	USD/VND	25,869.4 — with an honest note field: "DAILY REFERENCE rate, not real-time trading price"
gastracker	$0.001	1.8s	Base chain	Block 52,809,546, base fee 0.005 gwei, EIP-1559 slow/standard/fast tiers
macro	$0.002	1.7s	fed_rate	3.88% (2026-10-15), 12-day history, change +0
defiyield	$0.001	1.4s	aave pools	10 pools, top APY 1,123% (SDYFI), TVL + IL-risk flags included
polymarket	$0.006	2.2s	"bitcoin price end of october"	Live BTC markets: Yes 0.9995 / No 0.0005 on $72k, volume + liquidity per market
defi	$0.010	2.4s	USDC contract 0x8335…2913	Verified contract, 4.29B supply, 1,852 bytes code, risk flags [], "not financial advice" disclaimer

What the numbers tell us:

• Sub-2.5s across the board. Nothing here will bottleneck an agent loop. The slowest (defi, 2.4s) does on-chain verification work; the fastest (defiyield, 1.4s) is a cached aggregation.
• The honest note fields matter. fxrates tells you it's a daily reference rate, not a trading price. gastracker tells you tiers are heuristic multipliers. defi disclaims "heuristic signals only." We'd rather an agent know the data's limits than hallucinate confidence.
• Cost at agent scale: 1,000 news queries = $2.00. 1,000 DeFi yield checks = $1.00. 1,000 gas estimates = $1.00. For a research agent doing 100 queries/day, that's $0.10–$0.20/day — roughly the cost of the LLM tokens to read the results.

What we didn't measure (yet)

• Exa / Tavily head-to-head — Part 2. We'll publish the raw query log alongside scores so anyone can re-run it.
• Uptime under load — single-client measurements only. The public status page tracks ongoing availability.
• The other 34 endpoints — same method, future posts. tokenprice ($0.001) vs price-feed incumbents is next in the queue.

4. Transparent pricing: 10 core endpoints

No "contact sales." No tiers. Every endpoint has one price, paid per call in USDC on Base:

Endpoint	What it does	Price/call
websearch	Web search (Serper-backed), real-time results	$0.010
aiverdict	Multi-model AI verdict / decision scoring	$0.015
news	Crypto/tech news feed	$0.002
defi	Contract intelligence (verification, supply, risk flags)	$0.010
fxrates	FX rates (daily reference)	$0.004
gastracker	Gas estimates (EIP-1559 tiers)	$0.001
polymarket	Prediction market odds + volume	$0.006
macro	Macro indicators (Fed rate, history)	$0.002
defiyield	DeFi pool yields + TVL + risk flags	$0.001
sentiment	Sentiment analysis on text/social	$0.030

Plus 31 more (PDF tools, SEC filings, domain tools, GPU spot oracle at $0.10/call…) — same model, same transparency. If a price ever changes, the receipts page shows it; there's no invoice to argue about because every payment is a chain transaction.

5. The architecture: trust without accounts

Three design choices, all downstream of "agents shouldn't need accounts":

• Free trial, no signup. Per-IP daily quota by price tier: 10 calls/day for endpoints ≤$0.002, 5/day up to $0.01, 2/day up to $0.10, 1/day above that. No email, no key, no form. The X-Free-Remaining header tells your agent exactly how many it has left. If the data doesn't work for your use case, you've lost nothing but a few HTTP calls.
• On-chain settlement you can verify. Every paid call settles in USDC on Base. The transaction hash is the receipt — check it on a block explorer, not on our word.
• Public receipts page. Pricing, uptime, and recent settlement activity are public. When your agent spends $10 on 1,000 websearch calls, anyone can audit that the math holds.

This is the part we're proudest of, and the part most "AI API" companies won't copy: verifiability is a feature, and it costs us nothing to give away.

6. What's next + try it yourself

This series benchmarks each endpoint against its incumbent the same way: fixed queries, blind scoring, real cost-per-call math, raw logs published. Next up:

1. Part 2: websearch vs Exa vs Tavily — the head-to-head, with the full 25-query protocol above.
2. Part 3: tokenprice ($0.001) vs price-feed incumbents — a benchmark that costs less than a coffee to run.

If you're building agents and want to poke at the endpoints:

• Pricing & docs: https://api.mekongtradeintel.com/pricing
• Free trial: no signup — just call the endpoint and the trial quota applies automatically. Watch for the X-Free-Remaining header.

And if you re-run our baseline and get different numbers, tell us. A benchmark that can't survive reproduction is just an ad — we'd rather be wrong in public than right in private.