gpt-5.2 codex underperforms compared to gpt-5.2 for long-running coding tasks
What version of Codex is running?
0.81.0
What subscription do you have?
pro
Which model were you using?
gpt-5.2 codex
What platform is your computer?
_No response_
What terminal emulator and version are you using (if applicable)?
_No response_
What issue are you seeing?
i compared the gpt-5.2 codex model against the base gpt-5.2 model across many long-running coding tasks. the codex variant is advertised as being optimized for long-running complex coding tasks, but in practice i couldn’t find any advantage. for each task i ran the same prompts with gpt-5.2 codex and gpt-5.2, resetting the conversation each time. gpt-5.2 consistently produced higher-quality outputs with fewer bugs and better code quality. the codex model’s responses were often overly concise, missing key information during planning. in agent workflows where you plan before coding, a detailed plan is critical; codex’s lack of detail made the implementation phase a nightmare and degraded the resulting code quality. to work around this, i had to switch back to gpt-5.2 for planning and documentation, then back to the codex model for implementation, which is cumbersome via slash commands and easy to forget. overall there is no measurable improvement in runtime or quality for the codex model, and gpt-5.2 even runs longer and thinks more thoroughly. i don’t see any scenario where gpt-5.2 codex is better than the base model, so training and offering this branch seems pointless and increases cognitive load for users.
What steps can reproduce the bug?
- use gpt-5.2 codex xhigh in codex to implement a complex coding task.
- reset the conversation and run the same prompts with the base gpt-5.2 model with the same xhigh thinking.
- compare the outputs for completeness, bug count, code quality and detail in planning.
in every case i tried, gpt-5.2 produced more detailed plans and higher-quality code, while gpt-5.2 codex was too concise, missed important information and yielded lower-quality implementations. gpt-5.2 codex should at least match or exceed the base gpt-5.2 model on long-running coding tasks. it should provide detailed planning information to support agent workflows, generate high-quality code with fewer bugs, and eliminate the need to switch models for planning versus implementation. users shouldn’t have to trade off quality for conciseness or manage multiple models to get reliable results. the issue is poor performance and usability.
What is the expected behavior?
_No response_
Additional information
_No response_
5 Comments
Curious, what kind of prompts and tasks did you have GPT-5.2{-codex} work on?
just one example, i let both models to handle a swiftui issue, which is a pure ui issue, and with the same prompts, because i first let the gpt-5.2 codex model to implement the task, and then i reset the conversation and used exactly the same prompts and let the gpt-5.2 to do the same task again, and then the codex model failed the task because the result is not what i want. but the gpt-5.2 model finished the job and the result is exactly what i want. and i have tried probably 10 tasks like this in this week. and i can tell every single time the gpt-5.2 will either outperform the gpt-5.2 codex model or at least they both have the same equal quality result
Found the same, -codex is way faster, but loses track faster + produces worse code than the "regular" models. I've stayed on
gpt-5.2 xhighsince it became available, and it's the combo that seems to produce the best and highest quality code.Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the
/feedbackcommand. That will provide us with additional details including logs and conversation text.I had similar experience. Currently, 5.2 Codex (high) is barely usable compared to 5.2. The only upside is that it is much faster, but it has nothing in common with 5.2(high). So it is quite disappointing.