SWE-bench performance is similar to normal gpt-5, so it seems the main delta with `gpt-5-codex` is on code refactors (via internal refactor benchmark 33.9% -> 51.3%).

As someone who recently used Codex CLI (`gpt-5-high`) to do a relatively large refactor (multiple internal libs to dedicated packages), I kept running into bugs introduced when the model would delete a file and then rewrite in (missing crucial or important details). My approach would have been to just the copy the file over and then make package-specific changes, so maybe better tool calling is at play here.

Additionally, they claim the new model is more steerable (both with AGENTS.md and generally). In my experience, Codex CLI w/gpt-5 is already a lot more steerable than Claude Code, but any improvements are welcome!

[0]https://github.com/openai/codex/blob/main/codex-rs/core/gpt_...

[1]https://github.com/openai/codex/blob/main/codex-rs/core/prom...

Ditched my Claude code max sub for the ChatGPT pro $200 plan. So much faster, and not hit any limits yet.

Granted, it's not an apples-to-apples comparison since Codex has the advantage of working in a fully scaffolded codebase where it only has to paint by numbers, but my overall experience has been significantly better since switching.

For people that have not tried it in say ~1 month, give Codex CLI a try.

It is super annoying that it either vibe codes and just edits and use tools, or it has a plan mode, but no in-between where it asks me whether it's fine it does a or b.

I'm not understanding why it lacks such a capability, why in the world would I want to choose between having to copy paste the edits or auto accept them by default...

Both were struggling yesterday, with Claude being a bit ahead. Their biggest problems came with being "creative" (their solutions were pretty "stock"), and they had trouble making the simulation.

Tried the same problem on Codex today. The design it came up with still felt a bit lackluster, but it did _a lot_ better on the simulation.

I just don't see how this can be considered productive - I am waiting 20 minutes staring at it "thinking" while it does trivial tasks on a virtually bare repo. I guess for async agents its not such a big deal if they are slow as molasses, as you can run dozens of them, but you need a structured codebase for that - and i am already hours in and haven't even gotten a skeleton.

I have read through all the docs, watched the video. It would be so much quicker just to write this code myself. What am I doing wrong? Is it just super slow because they are over capacity or is this just the current state of the art?

I’m imagining if it can navigate the codebase and modify tests - like add new cases or break the tests by changing a few lines. This can actually verify if the tests were doing actual assertions and being useful.

Thorough reviewing like this probably benefits me the most - more than AI assisted development.

It sounds like it can make simple tasks much more correct. It's impressive to me. Today coding agent tends to pretend they're working hard by generating lots of unnecessary code. Hope it's true

And before someone says it, I do happen to have my own codex like environment complete with development containers, browser, github integration, etc.

And I'm happy to pay a mint for access to the best models.

Beyond that, purely anecdotal and subjective, but this model does seem to do extensive refactors with semi precise step-by-step guidance a bit faster (comparing GPT-5 Thinking (Medium) and GPT-5 Codex (Medium)), though adherence to prompts seems roughly equivalent between the two as of now. In any case, I really feel they should consider a more nuanced naming convention.

New Claude Sonnet 3.7 was a bit of a blunder, but overall, Anthropic has their marketing in tight order compared to OpenAI. Claude Code, Sonnet, Opus, those are great, clear differentiating names.

Codex meanwhile can mean anything from a service for code reviews with Github integration to a series of dedicated models going back to 2021.

Also, while I do enjoy the ChatGPT app integration for quick on-the-go work made easier with a Clicks keyboard, I am getting more annoyed by the drift between Codex VSCode, Codex Website and Codex in the ChatGPT mobile app. The Website has a very helpful Ask button, which can also be used to launch subtasks via prompts written by the model, but such a button is not present in the VSCode plugin, despite subtasks being something you can launch from the VSCode plugin if you have used Ask via the website first. Meanwhile, the iOS app has no Ask button and no sub task support and neither the app, nor VSCode plugin show remote work done beyond abbreviations, whereas the web page does show everything. Then there are the differences between local and remote via VSCode and the CLI, ... To people not using Codex, this must sound insane and barely understandable, but it seems that is the outcome of spreading yourself across so many fields. CLI, dedicated models, VSCode plugin, mobile app, code review, web page, some like Anthropic only work on one or two, others like Augment three, but no one else does that much, for better and worse.

I like using Codex, but it is a mess with such massive potential that needs a dedicated team lead whose only focus is to untangle this mess, before adding more features. Alternatively, maybe interview a few power user on their actual day to day experience, those that aren't just in one, but are using multiple or all parts of Codex. There is a lot of insight to be gained from someone who has an overview off the entire product stack, I think. Sending out a questionnaire to top users would be a good start, I'd definitely answer.

npm ERR! code 1 npm ERR! path /usr/local/lib/node_modules/@openai/codex/node_modules/@vscode/ripgrep npm ERR! command failed npm ERR! command sh -c node ./lib/postinstall.js npm ERR! /usr/local/lib/node_modules/@openai/codex/node_modules/@vscode/ripgrep/lib/download.js:199 npm ERR! zipFile?.close();

It is inevitable.