gpt-5-codex high is far worse than gpt-5 high on complex problems

It's probably using the same "router" as ChatGPT. I find that when using GPT-5 (auto) to ask questions, the answers are almost consistently wrong, but when using Thinking they're decent and with Pro usually quite good.

I've found gpt-5-codex to be quite dangerous to use. It behaves similar to some older Sonnet models, where sometimes when trying to solve a difficult problem, after being unable to solve it for a while, it starts doing all kinds of random shit like editing unrelated code, messing with git, deleting files etc. This should have triggered longer thinking instead, but seems it doesn't. End up having to spend as much time cleaning up after it as doing actual coding.

Codexgpt5 has a lot of problems, but it is also quite smart, but it is overly complex. A small problem can be solved in an hour, and it cannot trigger automatic compression. It really makes people love and hate it.

I believe gpt5codex also uses a routing mechanism for short thinking on simple tasks, but this routing mechanism doesn't seem to perform very well from the web interface's perspective. When I use the auto model on the web, I usually need to add prompt words like "think harder" to my prompt, which can route it to the thinking model, and the answer quality is like night and day. I think I could probably do the same thing with gpt5codex to see what the effect is. In that case, what's the point of gpt5codex's low, middle, and high settings? (sam: oh, don't ask me :))

It's probably using the same "router" as ChatGPT. I find that when using GPT-5 (auto) to ask questions, the answers are almost consistently wrong, but when using Thinking they're decent and with Pro usually quite good. I've found gpt-5-codex to be quite dangerous to use. It behaves similar to some older Sonnet models, where sometimes when trying to solve a difficult problem, after being unable to solve it for a while, it starts doing all kinds of random shit like editing unrelated code, messing with git, deleting files etc. This should have triggered longer thinking instead, but seems it doesn't. End up having to spend as much time cleaning up after it as doing actual coding.

I'm noticing the same exact thing. Especially with large monorepo setups.

I had it push the code, and after the push was completed, it performed a series of completely meaningless operations.
<img width="910" height="2027" alt="Image" src="https://github.com/user-attachments/assets/abe2e855-c3aa-4a59-925c-d956f9ca7ad3" />

Don't have any direct logs to share atm, but have had a lot of difficulty with the codex models since release. They seems to get stuck more often running commands, enter into command loops they cannot get out of, wait significantly longer for code diffs to be generated (because codex model is stuck running seemingly useless commands). It seems that reading a file can range from being a trivial task to something that makes the agent fall flat on its face. Right now going back to gpt-5-high has restored my sanity.

Example prompt I am giving: Grep the following files, build up an understanding of what the files are doing, based on this understanding, refactor the code to achieve xyz. (So usually reading multiple files, advising it to explore related files, building up context, and then executing a refactor that might span 5- 10 files max).

(Currently operating in monorepo environment)

The Codex models definitely "feel" more tuned to be used in agentic tools, but not in a good way.

The beauty of GPT-5 (especially gpt-5-high) in Codex CLI is that it truly excels in reasoning and stuffing itself with as much context as possible, regardless of the task.

Every other model before GPT-5 gave off this feeling of "I just got access to this codebase, I have no idea what’s going on." and it seems that no amount of initial prompting or worthless MCP tools solved this issue. However, GPT-5 does a great job of thoroughly navigating a repo before giving its first response, which in turn increases the quality of its next actions.

This probably belongs to Discussions since it's regarding a model and not the CLI/extension.

@dedrisian-oai Marking this as completed is a bad joke. This is a real problem and not fixed at all. Please revert that status ASAP.

@dedrisian-oai Plus, here are 32 likes on that bug issue. Closing this as "completed" without any fix is wildly unprofessional!

@dedrisian-oai It's your reponsibility to investigate or make the right people aware of this issue, or at the very least keep the issue open, not to push bad news under the rug. Pushing bugs under the rug certainly isn't going to help the product.

@tibo-openai please take a look and if necessary discuss with @dedrisian-oai when it is or isn't appropriate to close a bug as "completed". thanks for not pushing real bugs under the rug.

Tbh, i have a feeling that since codex with gpt5 that it is trying a lot of new things that for sure but it doesn't acknowledge or respect the codebase style anymore.

For instance if i would have like 20 cronjob with a query and a transaction event and ask him to do another one it would probably succeed partially in the task but not looking at the other cronjob as it used to do like before and replicate the transaction structure. Instead it create one of its own (that often doesn't work)

And for that i regret the old codex with gpt 4. Right now i fall in endless loop of code review to ask him to rectify something to respect the codebase and codestyle as for readabiility and KISS norms and then i discover that you have sometime to ask him to do something 2 time to him to fix it (otherwise it blatantly ignore you).

I probably won't be the last developer writing about that. so let's see the comments in the future. Might be quite interesting.

1) I've sent this feedback to the team to consider for model performance and differences between gpt-5-codex and gpt-5. Most of the feedback we have and benchmarks show that gpt-5-codex on medium is an excellent choice and appropriate for most, but your mileage may very.

2) Regarding the UX suggestion of allowing thinking times or budgets, I've noted this down and we'll consider, but this is not something we have planned for now.

Appreciate the suggestions and feedback. Will now lock this issue to avoid spam.