Hi all,
Feel free to share any thoughts about the gpt-5-codex model on Cursor here.
If you face any specific bugs, please create a Bug Reports
If GPT-5 Codex is only exposed via the Response API, what impact does that have on the backend? Are major pipeline/orchestration changes needed or is it a simple integration? And roughly how long until a stable release?
3 Likes
Thank you so much, Cursor team. About to test gpt-5-codex today and will come back here with thoughts.
2 Likes
The codex is so bad that any model is infinitely better.
After spending 4.51$ on Codex, I’ve had enough.
I am really sad. Had high hopes. Now I am switching to other model until Cursor tunes it’s interface with Codex. It just struggles to edit files.
Also I told Codex to create a new project besides my existing one and don’t touch the existing files. As you may expect, it anyways dug in.
Best model ATM. Just need some fixes on calling the edit file tool and sometimes it leaves the file with some issues.
Last time I tried it it was unusable. Fell back to codex cli for gpt-5-codex, gpt-5-high-fast in Cursor.
It seems to struggle with understanding the IDE tools. It has hallucinated edits after writing them in the chat. You may have to remove most tools from codex and just let it use Bash. It’s likely the outcome of the RL done by OpenAI on the model. It’s not a model for tool use, adding tools will likely give it worse performance even if they work.
1 Like
Still does the same thing gpt5 does to me, after it fills the context it just plain stops working. I really like the gpt5 models but this makes it almost unusable. I have to start a new chat for every request and if that request is large it fills context and stops working before it finishes ugh
For issues with gpt-5-codex, could you post a full separate Bug Report with more info Create Bug Report, perhaps also link them here so others can benefit.
Also add there also a Request ID with privacy disabled so we can look into the details: Getting a Request ID | Cursor Docs
It would help us adjust those cases we have not seen yet or if there are more common issues pop up.
hi @Kazopl we have it integrated already. It is already usable but as with every new model there might be tweaks by the AI provider and if we receive bug reports also adjustments from our side are possible.
1 Like
I think the model is worse than gpt-5-high.
1 Like
It struggles with the edit tool. Also if you accept changes while it is running, it loses its mind and starts deleting and reapplying changes it already made.
3 Likes
Hi, is it possible to share any request ids so we can diagnose the issue? 
Hi everyone!
Unfortunately, the integration of gpt‑5‑codex with Cursor has turned out to be pretty disappointing. This isn’t just my impression—there are already plenty of similar reviews out there.
When gpt‑5 first came out, it was free to try, so even with its flaws, it was still possible to use it productively. gpt‑5‑codex, however, feels different: there’s no promo period, and the integration itself is quite poor. Honestly, what’s the point of paying just to test something in this state?
I’m pretty sure it will eventually play nicely with gpt‑5‑codex and at least reach the level of Claude Sonnet 4. But right now, I can’t recommend it to anyone—better wait at least a month before giving it a try.
3 Likes
What people are saying here ??? gpt(-codex is a really really good model. You need to learn how to use prompt and when use reeasoning modele. CLaude 4 / 4.1 which was my prefered is so bad since 10 days approx. I completely stop to use and thanks you anthropic for that. The opportunity to test gpt5-codex was great and I love this modele it produice fewer ans fewer interaction and is really apeer coding assistant. Thanks to add it Cursor. I wanted to say Cursor team please refund all the $$ I spend with Claude 4.1 it was so so expensive for the quality produced in comparison to gpt5-codex. My favorite now.
I keep getting asked this but it doesn’t look there is a way to get a request ID that is useful to you without disabling privacy mode for our entire org. I don’t think I can do that even temporarily. If you can get individuals a privacy mode toggle, I would be glad to sandbox my repros and send you request IDs for these bugs I’ve been filing.
when will this gpt-5-codex model be available on the CLI, I quite like the CLI but there are quite a few models
oh….It used up 4.5 million tokens from me. Then it deleted all the generated files.
d0574a3b-cf75-436e-85d0-d43efa16de8a
I kindly ask - please add gpt-5-codex-high to cursor cli