Chat responses are consistently much shallower than ChatGPT Web with the same prompt and reasoning settings

Open 💬 0 comments Opened Aug 12, 2026 by metapak

What version of the Codex App are you using (From “About Codex” dialog)?

Chatgpt Pro

What subscription do you have?

Chatgpt Pro 20x

What platform is your computer?

_No response_

What issue are you seeing?

Additional information

This report is specifically about a systematic Desktop vs Web discrepancy, not normal stochastic variation between model responses.

I have already tested the same prompts across both clients, and the difference is consistently large enough to affect practical usability.

I am also aware that response latency alone is not sufficient evidence of a model downgrade. My concern is based on the combination of:

repeated identical-prompt comparisons,
materially reduced analysis depth,
consistently poorer task completion in Desktop,
and the fact that similar reasoning/configuration propagation bugs have previously been documented in Codex Desktop.

For that reason, I believe the most useful investigation would be to compare the server-side telemetry of an affected Desktop turn with the equivalent Web turn, especially the requested model, served model, requested reasoning effort, effective reasoning configuration, and any routing/fallback metadata.

What steps can reproduce the bug?

Feedback ID: no-active-thread-019ff45d-7f3d-7563-b66c-e30eab9599cd

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗