Vellium & CoTesting lab

AI agents now pay for data by the call. Almost nobody checks whether it's right.

We buy the same calls the agents buy, check each answer against an independent record, and publish the results for builders and sellers to see.

Our daily check, 11 October 2026 (still running)
14,378
paid data services checked so far, from the public x402 directory
98.4%
answered with a valid request for payment
189 ms
median time to answer
Coming soon
accuracy tests, with real paid calls

According to the directory's own records, agents made about 896,000 paid calls to the services checked so far in the last 30 days, worth roughly $64,000 at listed prices. Half of these services charge less than a cent per call.

Almost every service answers. Whether the answer is right is the open question, and it is the one we test.

Scorecard by area

Where an official record exists, we check answers against it. Where none does, we compare sellers with each other and over time.

AreaCheckedBought by 2+ agentsAnsweringChecked againstAccuracy
Crypto prices1,13188997.7%Exchange consensusComing soon
Company facts and filings47634297.3%SEC EDGAR and company registersComing soon
Stock prices52241296.7%Official closing pricesComing soon
Blockchain reads1,3471,05897.9%The chain itselfComing soon
Web search1,8751,37998.5%Freshness and agreementNext
Company and people enrichment90061499.0%Public registersNext
AI model answers86770498.3%Is it the model it claims?Next
News30326195.7%Agreement between sellersLater
Weather and places52341098.5%National weather recordsLater

Bought by 2+ agents: at least two different paying buyers in the last 30 days, from the directory's own records. Answering: replied with a valid request for payment when we called. One service can sit in more than one area.

How we test

The same four steps for every service, every day.

  1. Buy like an agentWe pay for each call the way any agent would, in USDC on Base, so we see exactly what a paying customer gets.
  2. Check against the recordEach answer is compared with an independent source: exchange prices, official filings, or the blockchain itself.
  3. Score over timeOne good answer proves little. Scores build from repeated checks, so a service that drifts or goes stale shows it.
  4. Publish with evidenceEvery score links to the calls we made and the record we checked them against.

For agent builders

Choose the services your agents pay for on evidence, before they spend.

The scorecard will come in a form agents can read directly, so they can check a seller before they pay.

For sellers

Ratings come only from our tests. No seller can pay for a better score.

You can see the evidence behind your score and ask for a re-test once you've fixed a problem.

About

This is the testing lab of Vellium & Co Ltd, an independent research firm. We don't sell data, and we hold no stake in any service we test.

info@vellium.co.uk