Auto Mode Became Practically Useless After the Recent Update

Ideas Feedback
1

Hey guys,

I don’t know what exactly you changed in Auto Mode recently, but it’s now only using the ChatGPT Codecs model - and honestly, it’s impossible to work with.

Previously, Auto Mode was running on Claude Sonnet 4.5. Sure, that model sometimes produced messy code, but at least it was fast, you could discuss the plan with it, and through long dialogues it could eventually arrive at the right solution - clean, documented, and consistent.

Now with Codex, it’s a total disaster. It takes ages to think, it ignores communication guidelines after the first summary, and constantly loops or forgets instructions. Sometimes it completes all requests, sometimes just one, sometimes adds random stuff, and still reports that everything is done.

So what I get now is:

  • Drastically longer response times,

  • Completely unpredictable behavior,

  • And a huge level of frustration.

The whole point of Auto Mode was that it should select the best model for the task. According to your documentation, Ultra Plan users should have access to all available models - and Auto should automatically pick the most suitable one.

But no matter how many new chats I start, it always defaults to this dumb Codecs agent. Auto Mode has become unusable. Please do something about it - either revert to the previous setup or give us back a model that at least communicates consistently, even if with a smaller context window.

Right now, it’s just awful.

1 Like
Auto Model suddenly using way more of my usage limit?
2

You could guess why I’m here in the first place. The recent update has been unusable in Auto. I closed down my project, waiting for the next cycle of subscription to renew before I can get back to it and select a model. Auto defaults to GPT 5 Codex and might be the worst experience I’ve had with cursor. Simple tasks can’t be completed, and the model repeatedly argues back over straightforward instructions. After repeating the same instructions many times, it will eventually execute them. This behavior needs to be fixed. The current auto mode is simply unusable for several reasons.

3

I just used auto a few minutes ago for explaining some code to me, and wrote a PR description based on what changes I have made - It did a good job.

I do have a couple of user rules set instructing all my models to answer in short form, do not hallucinate, and stuff like that. Maybe that’s giving me the benefit.