Do not judge a relay by one request. Test short connectivity, sustained requests, long-task recovery, billing behavior, and provider rules.
Short test
JuCodex: 100% availability, 1688 ms average latency, current status up.
YunDou: 100% availability, 1973 ms average latency, current status up.
WorldBase.ai: 99.9% availability, 2483 ms average latency, current status up.
Sustained test
Run 10 to 20 requests and at least one real coding task before increasing budget.
Record model names, errors, latency feel, and billing changes.