API Compass is live: why we keep a relay service's claims apart from our own measurements

The short version

This site does one thing: lay out what is known about AI API relay services, and mark clearly who said each piece of it.

That sounds simple, but most relay directories do not do it. The "9.9 rating" and the "0.05 rate" on their leaderboards are usually just lines the service wrote on its own promo page, copied into a table that looks like an independent verdict.

How bad is the information asymmetry here

You want to use Claude Code, so you search. You get dozens or hundreds of relay services, each one promising:

"Direct official channel, full-strength model, never downgraded"
"GPT rates as low as 0.05 — the lowest anywhere"
"99.99% uptime, three years in operation"

These claims share two properties: the service publishes them itself, and none of them can be verified.

And the decision in front of you is to pay money up front. That creates two structural risks:

The service disappears with your balance. Prepaid credit is a pool of money by design. A small operator can collect enough and vanish.
The model gets downgraded. You ask for Opus and get a cheaper model back. You will not notice unless you specifically test for it.

Neither risk is solved by "reading the ratings". Only one thing solves them: continuous probing plus a record you can look back at.

So we split the data in two

Field	Where it comes from	Can it be sorted?
Uptime, latency	Our own daily probes	Yes, sorted directly
Model count, model list	Probing the service's `/v1/models` endpoint	Yes, sorted directly
Endpoint verification status	Our own check	Yes
Rate, minimum top-up, promotions	Published by the service itself	Shown side by side only, with the source labelled

The last row is the point. A rate is a price the service sets. We are in no position to "measure" it — we would have to buy credit and test it, or simply reprint what they say. Since reprinting is the only realistic option, readers have to be told that is what it is.

That is why every table on this site that involves rates, minimum top-ups or promotions carries the line "advertised by the service; actual billing governs", and why those numbers are never sorted in the same table as our measured data.

Listing and ranking rules

Basic listing is free and you can submit at any time.
Ranking uses measured data only: composite score, uptime, latency, model coverage. We do not sell ranking.
There are ad slots, billed monthly, labelled "Promoted" on the page and kept physically separate from the data tables.
Services that stay unreachable, or that are shown to have exited with users' money or to have misrepresented their models, are removed, and the record of that is kept.

How to use this site

Filter out the unstable ones first using the rankings. Anything below 95% uptime over a sustained period drops out.
Then size the cost using the price list. Put your real usage into the rate calculator.
Finally pick one or two candidates and top up a small amount to test. No directory's conclusion replaces your own testing.

We will not tell you "which one is the best buy" — that depends on your workload, your volume and your appetite for risk. We only make sure that before you decide, you are looking at numbers that are as close to real as we can make them.