Do not judge a relay by one request. Test short connectivity, sustained requests, long-task recovery, billing behavior, and provider rules.
Short test
YunDou: 100% availability, 1967 ms average latency, current status up.
WorldBase.ai: 99.4% availability, 2780 ms average latency, current status up.
DBAI: 99.4% availability, 2819 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.