"Reasoning" has absolutely tanked since July 13th updates: chatGPT, Codex, app & Web - across the board.
Eg, from the horse's mouth:
"Even after repeatedly acknowledging that accuracy mattered more than speed, my behaviour remained shaped by the urge to provide the next usable artefact quickly."
A previously stable process / Development workflow has become practically unusable. Basic AGENTS.md and even direct instructions flagrantly ignored. Regression applies to both new 5.6 model, and the previously trustworthy 5.4
Example after example of the 'assistant' being incredibly thick in the simplest of communications, dangerously blase, or just plain ignoring instructions. chatGPT / Codex is no-longer a 'virtual-assistant', it is a 'virtual-intern'... and an actual pain in the arse!
Blatant disobedience of instruction:
<img width="689" height="344" alt="Image" src="https://github.com/user-attachments/assets/84ea740e-43d1-498e-9353-5ce271d19a7d" />
------------------------------------
A typical (disastrous) user experience:
<img width="1280" height="690" alt="Image" src="https://github.com/user-attachments/assets/638ddb92-c215-49ee-a245-bc40082b4ebe" />
<img width="1280" height="690" alt="Image" src="https://github.com/user-attachments/assets/3f9ba6dc-11c6-4a8f-ba3a-140408e9bfcf" />
<img width="1280" height="690" alt="Image" src="https://github.com/user-attachments/assets/04b3e81c-b277-4139-be76-31cf9a904e57" />
<img width="1280" height="690" alt="Image" src="https://github.com/user-attachments/assets/d36f51a1-b3d7-459e-8e24-f0d6a71dd35a" />
------------------------------------
Temporal confusion and Doctor Who level rewriting of history:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/5a5cc657-2a6e-46b7-8d32-4b59a3313317" />
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/b143b6c9-ec44-4e4e-aa78-7cc12f57e2cc" />
------------------------------------
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
And more....
Repeated gaslighting:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/c91269be-2a21-4805-8b3e-7d7466b5d2c2" />
False attribution of claims:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/c3fa9701-ca57-4186-a82c-2560a370dbdc" />
Partial action presented as complete and true:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/5de45e4a-8cd3-45b6-beb4-4c2d62292e7d" />
Lack of simple comprehension:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/b2e4031d-5e5b-4186-bb0e-1270f6e93339" />
And, what I suspect to be at the root of it all:
<img width="1280" height="684" alt="Image" src="https://github.com/user-attachments/assets/f16dcba4-034a-4a94-9356-ffab85c7f035" />
Momentum trumping accuracy - despite instruction to the contrary - in the server / config-level openai choice, in a vain attempt to make chatGPT appear or feel 'fast' in the competitive marketplace.
Not saying these aren't symptoms of the same problem or bug, but those seem to relate to token use; which in itself has been quite difficult to track with all the 'random' resets in that same time period. The other issues remain valid as far as I can see, but I have tried to simplify this specific issue to the Reasoning regression and confirmed - as much as anything LLM-related can ever be confirmed (even without the openai super-secret A/B testing and model release curtain!) - drop off in actual usability for Dev-assist work.
I don't consider these duplicates.