Codex quality drop may be tied to the app, not the model
Switching from Codex App to another tool that can use the same Codex made the coding results much better on the same codebase and the same prompts. Before the switch, GPT on XHigh felt worse than a free Chinese model, but that problem went away after changing tools.
was not tested. Because the inputs stayed the same, the likely cause is the Codex App or its running environment, not necessarily the model itself.
Key points
- Changing away from Codex App appeared to fix the quality problem.
- The same Codex was still used.
- The same codebase and prompts were used for comparison.
- was not tested.
- The suspected problem is the app or running environment, not necessarily GPT itself.