I use AI API relays myself, and I kept hitting the same problem: a relay looks fine today — right model, low latency — and a few days later the model is swapped, or it falls over at peak hours. A single probe tells you almost nothing.
So I built hvoyai.com. It probes public AI API relays per model, over time, and publishes uptime, latency, pricing and sample size on public leaderboards, so you can see the long-run record, not one test.
For example, Opus 5.5 is listed by 106 relays, Sonnet 5.5 by 104, GPT-6.1 Sol by 113. Same model, very different price and stability across relays — the cheapest quote alone is a bad way to choose.
The site now covers 622 models from 77 publishers and 137 relays, updated daily. I don't sell API access and I don't operate a relay. I just measure. Take a look, and tell me in the comments what I'm missing.
