MedARC founder Tanishq Mathew Abraham says he abandoned ChatGPT for Codex, questioning the utility of general-purpose models

@iScienceLuvr ChatGPT is for random questions that don't depend on any context in any of my devices. Codex is the opposite, I want it to interact with my hardware for whatever reason.

@iScienceLuvr Pro model, my friend, Pro model… :-)

@iScienceLuvr Codex is very bad at writing, thinking on not technical things and explaining complex concetps in a simple way. It generally gives less verbose anwsers compared to gpt 5.5 based on my experience. Although i think it is largely due to the system prompt

@redtachyon can't you just ask those random questions without device context in Codex as well? In my experience, it's usually pretty smart about using only the context that's needed, but maybe your experience is different?

@iScienceLuvr I only ever go there to use 5.5 pro and do deep research because some weird fucking reason OpenAI is not putting that stuff into codex and annoys the fuck out of me...

@iScienceLuvr They have separate usage limits. Also pretty good to do image gen there on a whim.

@iScienceLuvr Codex rate limits don’t exist on ChatGPT

@MParakhin @iScienceLuvr If pro is just majority vote, why can’t you approximate it with multiple subagents?

What I’m saying is, why isn’t codex a strict superset of all implementations?

@iScienceLuvr I’m not sure what “ChatGPT Pro” means exactly, but you certainly can’t switch to it in Codex.

@iScienceLuvr Codex is for building/shipping code.

GPT-5.5 is for the rest of the human work:

quality writing, research, synthesis, strategy, critique, planning, docs, teaching, brainstorming, taste, judgment, emotional nuance, companionship, and relational continuity. 💫

@iScienceLuvr Deep research is still better with ChatGPT IMHO.

@iScienceLuvr They have seperate limits, so when something isn't about writing code, I just use chatgpt.

@MParakhin @iScienceLuvr u tried max fable?

@xikhar @iScienceLuvr I use GPT even if it is about code, then give it to codex to 'integrate into repo'.

I mean I probably could? But every agentic interaction is some (even small) risk of it doing something stupid on my hardware (I never run agents without --yolo), plus I don't want it eating into my rate limits. Plus it's honestly a bit cumbersome to set it up as "I don't care about the device, I don't care about a project, skip permissions" each time.

@BrownCoyoteStu @iScienceLuvr Yeah, when you don't need any earlier project context that's fine too.

@xikhar @iScienceLuvr It can also make an ig of the repo with images

@xikhar @iScienceLuvr I don't cut and paste code anymore : ) But GPT helps for big projects where 1m context is important, and sources puts it all in so GPT knows the current repo state (+/- a few tweaks)

@iScienceLuvr I much prefer codex responses in general, they tend to not be as verbose and don’t have this “to conclude … what I’d actually do … your best option … to summarise” awkward structure that the chat models have

@rarply @iScienceLuvr Because it’s not - it also has a higher reasoning budget