Knowledge base · TOPIC-03

ChatGPT works, but Codex fails

Short answer: The two product surfaces can exercise different destinations, protocols, process contexts, and connection lifetimes. ChatGPT working is useful evidence, but it is not an end-to-end test of Codex.

OpenAI documents secure WebSocket requirements for ChatGPT and Codex and lists product-specific destinations. That statement is provider-reported. Your browser page load is an observation of that browser path. The claim that the two paths differ in your specific case remains inferred until both are measured.

What the comparison tells you

If ChatGPT works, general Internet access and at least one OpenAI application path are available. It does not establish that Codex resolved the same destination, chose the same IPv4/IPv6 path, inherited the same proxy policy, trusted the same certificate chain, or maintained the same WebSocket session.

Safe checks

• Record the failing Codex action and time.
• Check official component status without treating it as local-path proof.
• Compare DNS, TCP, TLS, HTTPS, and—only at approved destinations—WebSocket evidence.
• Keep authentication separate: never paste tokens or cookies into a diagnostic tool.
• Change one variable by trying one trusted alternate network, then repeat the same Codex action.

The correct result may be “ChatGPT browser path observed; Codex native path unknown.” Open Probe V0.1 can collect only its supported OpenAI DNS, TCP, certificate-validated TLS, governed anonymous HTTPS, and repeated short-window evidence. It does not measure Codex WebSocket/session continuity or proxy inheritance. Recovery must still be verified by completing the same Codex workflow.