OpenAI/Codex usage needs sustained stability, not just one successful request. Coding, file reading, long-context work, and tool loops amplify relay instability.
Start small with monitored candidates such as YunDou, lingxicode, SudoCode, but do not treat ranking as the final answer.
Testing steps
Run a real 20 to 30 minute task and watch for 429, 5xx, timeouts, truncation, unavailable models, and latency spikes.
For high-frequency OpenAI/Codex usage, keep one primary relay and one backup relay.
Candidates
YunDou: 100% availability, 1995 ms average latency, current status up.
lingxicode: 99.4% availability, 2246 ms average latency, current status up.
SudoCode: 99.2% availability, 2235 ms average latency, current status up.
infai: 99.2% availability, 2311 ms average latency, current status up.